AI Use Cases
Real use cases people shared on Reddit, Hacker News, X, GitHub and blogs, boiled down to plain words with a link to every original post. Fresh finds land every day.
Showing 471 of 471
UK estate agents are using AI to enhance listing photos, from repainting walls and decluttering rooms to adding furniture or lawns, saying it helps buyers see a home's potential. Some listings went too far and were taken down, and one agent now labels every AI-enhanced photo.
Source: Web ·
A patient with Crohn's disease was refused his usual drug and put on a substitute that made him ill, so he used the AI appeal tool Claimable. It drafted a letter built on his symptoms, missed work and cited federal rules, and five days later his doctor's office called to say the original drug was approved.
Source: Web ·
An author built the Foldable Tractatus, an open-source web app for reading Wittgenstein's first book, using only $20 a month Claude and ChatGPT plans. They moved from Claude chat to Claude Design to Claude Code and a little Codex, never read the model-written code, and published all 191 prompts with an analysis.
Source: Blog ·
Inditex Tech built Kumoss, an open-source platform where LLM agents turn natural-language requests into infrastructure changes, including edits to resources already running. Validation, compliance checks and the usual review steps stay in the loop before anything is deployed.
Source: Medium ·
A hobbyist whose $300 embroidery machine wanted a subscription built a free open-source app instead. It turns SVG artwork into stitches, and Claude Code or Codex can draw and edit designs inside it through MCP.
Source: Hacker News ·
A company of about 20 people started with an AI bug triager for Sentry alerts and it grew into a system used for reporting and fixing bugs, code review, metrics and answering all kinds of questions. People reach it from Slack and the company forum.
Source: Hacker News ·
A reader working through a book on building a large language model from scratch photographs pages and chats with Claude to ask questions and check their understanding. They say they never stay stuck because there is always someone to ask.
Source: Hacker News ·
One builder made Cinnamon for their own family because most agents focus on personal tasks, and wanted both personal and shared contexts with easy-to-manage memories and facts. It runs on Gemini models entirely on Google Cloud and can share task results such as a marketing plan or a playable HTML game.
Source: Hacker News ·
A user set up OpenAI's dot, an always-on agent with its own browser that is connected to their calendar, email, Shipt and Granola. It unsubscribed them from about 50 mailing lists, proposed a week of simple dinners and built the grocery cart, and found four restaurant options, with a rule that it cannot send or delete email without approval.
Source: DEV ·
A parent built a free web app where Claude (Opus 5.5) reads a photo of loose LEGO bricks and designs builds from only those pieces, with 3D steps. A plain stud checker reports problems and Claude gets up to three rounds to fix them; each build takes a few minutes and costs about a couple of dollars in API calls.
Source: DEV ·
A creator split a 20.7-second TikTok ad for a portable blender into 14 shots, used gpt-image-2 to edit the keyframes and Seedance 2.0 mini to animate them, swapping the cup, table, milk and fruit. It cost $1.55 and about 10.6 minutes, and 8 of 10 spot-checked frames matched the request.
Source: DEV ·
A founder gave Claude Code a JSON record for each launch directory plus a browser driven by Playwright MCP, and it submitted one product to 285 of 338 directories. At last count 54 were live and most of the rest were in review, and the agent hands back forms it cannot finish.
Source: DEV ·
An AWS engineering team had agents write tests, find issues and fix them behind a merge gate, and says the team wrote and reviewed none of the code. Kiro produced a first Amazon SQS in about 15 minutes; the article admits gaps and says turning it into a robust product took much more than a weekend.
Source: Web ·
A traveller who is not an archaeologist had 48 AI agents review about 1,200 square kilometres of public lidar and flag a hilltop with a plaza and a pyramid. Mexico's heritage institute replied that it had no registered site there, and he later learned the site was already mapped in a 2025 study.
Source: Hacker News ·
A local maker used the open-source Qwen-Image model to redraw every block of three neighbouring towns in a city-builder style, working from Google's 3D map imagery. The page adds a sky that follows real weather and live plane positions, and says small details are invented.
Source: Hacker News ·
A traveller with no archaeology background brainstormed with Gemini, ChatGPT, Claude and Grok, then had 48 agents review about 1,200 square km of public lidar images overnight. One hilltop looked like an unrecorded Maya site, and Mexico's heritage institute replied within days that it had no registered site at that location.
Source: Hacker News ·
Someone managing type 1 diabetes trained and compared many glucose forecasters, including their own small transformer that predicts two hours ahead, on their own data and a public dataset, and runs the models on a phone. The author says it is research only and must not be used for treatment decisions.
Source: Hacker News ·
A developer built an add-on for Claude Code that, while you work as usual, has Claude write a native-sounding version of each English prompt with the reasons into a local file. A pane lists the corrections and can review recurring patterns in your mistakes.
Source: Hacker News ·
A lecturer's team found ordinary speech to text garbled English technical terms spoken inside Cantonese in live lectures. Their classroom tool now uses Gemini models in the cloud, or Gemma 4 E2B running in the browser, to produce live subtitles in nine languages.
Source: DEV ·
A team built an inbox assistant that runs fully on a laptop with a 9 billion parameter Qwen model in Ollama. It can read mail and calendar but cannot send, delete or forward. Sorting one email per request found all 16 emails needing action in a 37 email test inbox, and plain code checks flagged all 6 planted scam emails.
Source: DEV ·
A solo app developer who pays for Claude tried about 30 skills and kept five. One drafts App Store screenshot text in ten minutes instead of an afternoon, and a review skill caught a timer that kept running after its window closed.
Source: DEV ·
An engineer moved scheduled Claude routines into a git repo that six teams now use. In one month the agents helped close 73 escalation tickets and 10 accessibility issues and proactively fixed 5 bugs, opening draft pull requests that people review each morning.
Source: DEV ·
A Hacker News reader stuck on the finer details of a new family project started a voice conversation with ChatGPT on a morning walk instead of typing. It brought up angles they had not considered, they finished the walk in good time and came back more motivated.
Source: Hacker News ·
A solo builder of a Dutch health insurance comparison site has Claude extract structured data from 607 insurer documents, then validates it, diffs it against last year and has a person review every change: 8,540 items so far, with 1,407 rejected. AI assistants can then fetch the current numbers with sources instead of quoting last year's deductible.
Source: DEV ·
The maker of a small AI song generator in 22 languages set a daily cron job that, every three days, starts Claude Code headless to add the next language from a queue, deploy it, check it and email the result. The first fully unattended run added Norwegian.
Source: DEV ·
A lawyer runs DeepSeek V4.1 Flash on a self-built server with eight used Nvidia V100 GPUs for first-pass work on case files: it reads a folder of records, pulls out facts, dates and numbers, summarizes them and drafts routine text. Client files stay off the cloud, and nothing goes out without a real review.
Source: Web ·
The creator of a plotting library with a backlog of hundreds of issues and dozens of pull requests wrote guidelines matching what they care about in a PR and handed the project to Claude. Every PR is now merged or declined, issues are triaged and bugs fixed where possible, and spot checks found the code reasonable.
Source: Hacker News ·
An open project uses Claude Opus 5.5 to transcribe and deep-read digitized manuscripts by mathematicians such as Gauss and Riemann, about 12,000 pages so far at roughly $0.25 a page. Every candidate find goes through human reading, code checks and a literature search, and one lead on Riemann's shock-wave sheets is waiting for a palaeographer.
Source: GitHub ·
A chess player built a desktop coach that downloads their Chess.com and Lichess games, rates every move with Stockfish and has Gemini explain each one, with Gemini told to ask the engine about any what-if move. The app checks every move Gemini names, and the Gemini free tier works.
Source: GitHub ·
A developer set up a leaderboard repo and had 10+ Astra, Fable and ChatGPT agents keep trying to beat the current best Catan bot, merging in any version that won over 55% of games. The small CPU-only bot reached number 1 on Colonist's global 1v1 leaderboard with a 2552 rating.
Source: Web ·
A developer built an offline field guide where any phone on the same Wi-Fi sends a photo to a PC running the small open model SmolVLM2, which guesses the plant, bug or mushroom and looks it up in about 19,000 offline Wikipedia articles. Every guess carries a warning, since it once called a Christmas cactus a rose.
Source: DEV ·
While renovating a new home, someone who found only praise for appliances online wrote an agent skill for Claude Code, Cursor or Codex that looks up administrative penalties, inspection failures and recall notices, digs up scattered negative reviews and grades the evidence. If no option is reliable, it says none right now instead of pushing a product.
Source: GitHub ·
A developer worn out by juggling agents moved all work into markdown tickets in git: Claude Code and Codex research and build the tickets, and a Claude artifact page shows what changed so the developer can rule on one item at a time. They call it a big step forward in their productivity, though still tiring.
Source: Blog ·
A developer who did not want to pay for a cloud dictation service forked the open-source Handy app so the local Parakeet speech model reads words from the front window and uses them to settle close calls. Errors on on-screen words dropped from 5.83% to 3.81% on real slides, with nothing sent to the cloud.
Source: Blog ·
A Linux user handed everyday system chores to Codex: it found two audio daemons fighting over their Bluetooth headphones, fixed a capture card driver bug after sleep, set up a tool that cured laptop throttling and upgraded their small web server. Now their headphones work and the laptop runs much better under load.
Source: Web ·
A developer used Google Antigravity to build a small voice-cloning tool on Gemini 3.8 Flash TTS, cloned their own voice and, with consent, a family member's, then made a monologue in seconds and a playful two-voice dialogue about their cats. Both voices then showed up as custom voices in Google AI Studio.
Source: DEV ·
A developer made a family tree that lives on their own computer: they type sentences in Italian or English and a local Gemma model running in Ollama pulls out names and relationships. Their own code checks the result and shows a preview, and nothing is saved until they confirm.
Source: DEV ·
A small Sydney agency, bigpickle, built a research, proposal and review agent trio that shares markdown client notes, running on a local Qwen model through Ollama. A full run for a real bakery took 16 minutes and the reviewer approved the quote on its first pass with one low finding.
Source: DEV ·
A developer runs AI Werewolf, a group chat game where one human plays with 5 to 15 bots powered by models such as DeepSeek, Gemini and Qwen, plus AI voices and pictures. Over 53 days it cost $337.21, about $6.36 a day, with a real game averaging $1.49.
Source: DEV ·
A developer built a home-run assistant you chat with on Telegram: send a meal photo and Claude (the default model) estimates calories and macros, forwarded bank texts become a categorised spending ledger, and it nags you to bed until you say good night. Everything goes into one SQLite file on your own machine.
Source: GitHub ·
Someone studying a power electronics textbook turned it into a local study web app: a model of their choice, such as DeepSeek or a local Ollama model, writes pre-reading questions, line-by-line explanations, practice problems with strict marking, and plays a beginner who quizzes you as you explain a topic back.
Source: GitHub ·
A developer's open-source tool built on Claude Code reads your CV, finds and ranks matching jobs, writes a tailored CV and cover letter for each, then fills in applications in a browser window you can watch. Nothing is submitted until you say yes, and it never invents experience.
Source: GitHub ·
A developer's app sends a photo of a grocery receipt to the Gemini API and gets back a list of items with quantities, prices and expiry dates. Once the user confirms the list, the app tracks what expires soon and suggests recipes to use it up before it spoils.
Source: DEV ·
A playful site turns a short sentence into speech with the Piper text to speech model and plays it reversed so you hear the target sounds. You then record yourself and play your attempt reversed to check how close you got.
Source: Hacker News ·
A developer built an email helper with the Gmail API and Google Gemini that works inside a Mattermost chat room. It lists recent emails, searches the inbox on request and drafts replies for review, but it never sends an email on its own.
Source: DEV ·
A developer wrote an open-source skill for Claude Code and Codex that picks companies matching an offer, checks who is responsible on official pages, keeps the sources and drafts outreach. In a five-company test it marked unclear roles and missing contacts for review instead of guessing, and found no personal email addresses.
Source: DEV ·
After a Hacker News question about AI coding setups drew replies that were almost impossible to sort through, the poster used Claude to pull the tools, models and plans out of each comment. The result is an interactive site mapping 750 comments, with every number linked back to the comments it came from.
Source: Hacker News ·
An open-source project hands an AI agent the LDraw source file of a LEGO model. The agent infers its mechanisms, plans how the parts interact and writes Blender Python code that rigs and animates them; the maker's runs used Claude Sonnet and Opus as the agent.
Source: Hacker News ·
A student's free iOS app for recording shrine visits and goshuin stamps added a button that asks Gemini for a short introduction to the shrine, told by a companion deity character. The text is only generated when the user presses the button, and visit records stay on the phone.
Source: Blog ·
A member of a small infrastructure team built a Claude Code skill that sends Mattermost direct messages for them. It reads a personal style guide first, so the agent's five-paragraph drafts become the one-line question they would have typed themselves.
Source: DEV ·
A local MCP server reads FreeStyle Libre glucose data from a LibreLinkUp follower account so an assistant such as Claude Desktop can answer questions about it. Every answer says how fresh and complete the data is, stale readings are labelled stale, and credentials stay in the computer's keychain.
Source: GitHub ·
A college student who found it hard to gather enough players built a word association game where any seat can be a person or an AI model, so one person can get the full strategic game alone. The site also rates models such as GPT-6 Astra and Kimi by win rate and clue accuracy.
Source: Hacker News ·
A tester at a software company built a multi-agent system that reads requirements, writes test cases, drives web and mobile apps and files defect reports, with a person confirming each bug before it goes to a developer. They schedule runs before leaving work, and on one project the agents found 21 of the 30 defects confirmed as valid.
Source: DEV ·
A solo Roblox developer chained the FLUX image model on Cloudflare Workers AI, Blender and the Roblox Open Cloud API, with an AI coding agent running steps inside Blender and Roblox Studio over MCP. They type a short description and a few minutes later a colored, game ready model is sitting in their Studio place.
Source: DEV ·
A one person studio runs 13 standing AI agents and records every item they handle in a handover ledger, which also feeds a weekly report. When ChatGPT started sending one page 515 visitors in a week, the agents added an email signup box and later moved it higher on mobile, but any email send waits for the owner's approval.
Source: DEV ·
A gamer who plays only on CachyOS Linux bought a Flydigi Apex 6 Pro controller and had the Hermes agent write a custom app for controlling its haptics. They run Hermes with several models, including a local Qwen quantized to fit their own GPU for easy jobs, and are also generating pixel art locally for a Godot game.
Source: Hacker News ·
A developer brainstorms ideas for apps for personal use in ChatGPT over a few prompts, then asks it for a prompt to build the app in OpenCode with the low cost Muse Spark 1.3 contributor model. They say it usually gets done in one shot or with a little refinement and costs only a few cents.
Source: Hacker News ·
A hackintosh hobbyist installed macOS 26 on a spare ThinkPad T490 and used the Claude Code and OpenAI Codex integrations in Xcode 26 to build their first iOS app. They had a working app within a day without writing a line of code themselves, after a few x86 workarounds to get Claude Code running.
Source: Blog ·
A person with cochlear implants in both ears, who could not get anyone to explain what they actually hear, had Claude build browser test apps and ran seven psychophysical tests on themselves, 10,992 trials in all. They published the trial data, the analysis code and Lean 4 checks of the apps' logic, and used Claude to translate the write up into English.
Source: Hacker News ·
A research project had five small local models (Llama 3.2 3B, Mistral 7B, Gemma 2 2B, Phi-3 mini and Qwen 2.5 3B) answer questions with citations through Ollama, and used Claude Sonnet 5.5 only to judge whether each cited passage backed the answer. On 1,000 questions their rule lifted accuracy from 77.1% with a simple majority vote to 85.0%, though Claude alone picked the right answer more often on the questions where the models split.
Source: Blog ·
A coding teacher wrote beginner Python explanations, embedded them with Voyage AI and had Claude answer questions only from the closest passages, with citations. The very first question, what a function is, revealed that their notes never actually defined one, so the assistant doubled as an audit of the teaching material.
Source: DEV ·
A developer built a survival kit that works with no internet on Windows, Linux or macOS: a small local model (Qwen 2.5 3B by default, through llama.cpp or Ollama) answers questions from bundled guides on water, fire, food and first aid in Portuguese, English and Spanish, citing its sources, next to Kiwix offline libraries.
Source: Hacker News ·
This open-source Linux tool notes which apps you used, how long you stayed focused, and how often you switched, without recording what you typed. At the end of the day a model, local Qwen by default, writes a short reflection on your main work sessions and distractions so you can plan tomorrow better.
Source: DEV ·
This video shows how to put a tiny story-writing model on an ESP32-S3 board and run it with no cloud at all. It goes from downloading the project and building the firmware to flashing the board and watching the first short story print out. A fun, low-cost way to see how small an AI model can be.
Source: YouTube ·
A Japanese learner had Gemma retell five public-domain ghost stories at all five JLPT levels, from very simple kana up to near the original Meiji-era text. You read one scene at a time, answer a question, tap any word for its reading and meaning, and export your new words as an Anki deck. The original text is always one tap away, and progress stays in your browser.
Source: DEV ·
When someone dies, their inbox is often the only record of fixed deposits, insurance, shares, and subscriptions. This open-source tool has a local Gemma model read the mailbox on the family's own laptop and list each item with the exact email line that proves it, while ignoring scam emails. Claim letters are drafted, but nothing is sent until the executor approves on their phone. It was tested on a hand-written made-up mailbox.
Source: DEV ·
Made for a freelancer sister, this free web app reads invoice PDFs and receipt photos, has a small Gemma model running on your own computer pull out the key fields, and exports a spreadsheet ready for your accountant. Client names never leave your machine, and it also predicts which unpaid invoices are likely to come in late.
Source: DEV ·
Answer four questions about someone you love, and Gemma writes a short Javanese wayang play about them, with plain code checking the script before your browser performs it with puppets, an oil lamp, and gamelan music. Each play costs a fraction of a cent and the story isn't stored. The maker also taught a small Qwen model to write the plays for under a dollar.
Source: DEV ·
Instead of hiring someone just to dig through logs, a team pointed an open-source SRE agent at their Grafana alerts. A small Qwen model on an old graphics card reads the logs, metrics, and Kubernetes state, then posts a summary card to the team chat a few minutes later. It's read-only, costs nothing per alert, and no data leaves their network.
Source: DEV ·
This project scans the sensor records of a European Space Agency satellite that burned up in 2024 and flags the signals that look wrong. A small Qwen model, fine-tuned for under a dollar, answers questions using only the mission documents, and when the cause isn't written down it says so instead of guessing.
Source: DEV ·
In a simple bar game, a research team secretly paired up two AI players and asked local models to spot the cheaters. The AI judges did little better than a coin flip, and once even missed a player saying "our plan" out loud, while a two-line check of who kept winning together scored about 0.9. A good reminder to try simple code before trusting an AI judge.
Source: DEV ·
Coding agents often create servers, databases, and deploys mid-task and then move on. This free, open-source command-line tool reads the logs Claude Code and Codex already keep on your computer, lists anything that might still be running and costing money, and gives you the exact command to shut it off. It only reads, it never changes anything.
Source: X ·
This free tool opens small games like Snake right next to Claude Code or Codex. When the agent finishes or needs your approval, the game pauses, beeps, and shows a banner, so you get pulled back right away instead of drifting off to another tab. No account and no tracking.
Source: Hacker News ·
Phone screen recordings don't show touches, so this developer works them out from the video itself. Simple pixel math in the browser finds when the screen reacts and whether it was a swipe or a tap, and only a pair of small before-and-after stills goes to Gemini to ask what was tapped. The best model found all five taps in a test clip for about two cents per ten taps.
Source: DEV ·
A developer built a text version of a classic tank game with Claude Code in an afternoon, then had four Claude models play 29 matches. The biggest, slowest-thinking model lost almost every time, while a steady mid-sized model kept winning with a plain, careful plan. Every prompt and reply is shared, so you can see how each one played.
Source: DEV ·
This small command-line tool runs Gemma on your laptop to ask one gentle question at a time about someone you want to remember. Your answers are saved word for word in a Markdown file, and the AI never writes the life story or fills in gaps, so nothing ends up in the keepsake that nobody said.
Source: DEV ·
In this short video, a developer makes dinner while chatting with Codex. Codex uses tools and the browser on his Mac and hands jobs to a coordinating agent on his Linux machine, which compares search models and tests an app. Results come back into the same conversation, and anything that needs permission waits for his OK.
Source: Blog ·
This open-source Chrome extension takes a quick screenshot of the page you're on and has Gemini look for dark patterns like fake timers, guilt-trip opt-out buttons, sneaked-in add-ons, and fees that appear at checkout. Anything suspicious gets a red box drawn around it, and the screenshots aren't stored.
Source: DEV ·
Over three weekends, a developer had Codex and Claude Code make a whole strategy game, including the code, sprites, animations, and music. Then the same agents were set loose to play it live, gathering resources, building units, and protecting their kings, to see how coding agents cope when they have to think in real time.
Source: X ·
This free skill for Claude Code (and other coding agents) runs a security review where every finding has to show the file and line, the exact request an attacker would send, and the fixed code. No proof, no finding. On a test shop API with 10 planted holes and 3 safe-looking traps, it caught all 10 and left the traps alone.
Source: DEV ·
Instead of giving Claude SSH access, this developer gives it a narrow open-source tool that reports on the apps running on a Raspberry Pi home server. Eleven days of testing showed the tool itself was guessing, so it was rewritten to say only what the machine actually reported. A good rule for any tool an agent reads: no plausible guesses.
Source: DEV ·
This small local lab shows how a secret can survive in an AI app's search chunks, summaries, and saved answers after the original document is deleted. It tracks what was made from what, so cleanup removes every copy, then runs checks to prove the secret is gone and unrelated facts still answer. Everything uses made-up data and a local Gemma model.
Source: DEV ·
Git finds commits that rewrote lines from an earlier fix, then a local Gemma model reads both and decides if the same bug came back, the fix caused a new bug, or it's unrelated. Plain checks keep it honest, and the maker could confirm only one of 14 flags, a good reminder to verify.
Source: DEV ·
Pakistan's central bank publishes the names on long-dormant accounts so families can claim the money, but it's 14 million rows across hundreds of spreadsheets. This free tool searches them all at once and handles spelling variations in English and Urdu, the matching runs in your browser so the name you type isn't sent off, and Gemma helps rank matches and drafts a claim letter. The bank still makes the final call.
Source: DEV ·
The maker of a football quiz site asked eight models which clubs 200 players had played for and scored every answer against a hand-checked list. Even the best one got about a quarter of its clubs wrong, so the quiz data stays checked by hand. A handy way to measure a model on your own facts before you let it write anything.
Source: DEV ·
Web pages can hide text people never see, like white-on-white words or tiny fonts, telling an AI assistant to lie or push a product. This open-source tool finds that hidden text and removes the harmful parts, as a browser extension, a page checker, or an API. In the maker's test, a model obeyed hidden orders on 19 of 20 pages, and on none after cleaning.
Source: DEV ·
A parent built a small game where his 10-year-old designs the rooms, ghosts, and clues, can draw a monster on paper and meet it in the game, and then types what to do in her own words. A local Gemma model tells what happens, while simple code keeps track of doors, items, and score so the AI can't bend the rules.
Source: DEV ·
This developer lets Claude Code write every fix and setup decision into an Obsidian vault, with a rules note saying where things go and that it must never delete his own text. Claude Code runs on a home server next to a synced copy, so he can ask his notes questions from the Claude phone app wherever he is.
Source: DEV ·
This free local guard hooks into Claude Code and checks every command before it runs: packages that don't exist, look-alike names, sketchy install scripts, and READMEs with hidden notes for AI. It blocks only the risky part and tells the agent why, then a local Gemma model explains the block to you in plain English.
Source: DEV ·
Four Claude models sit in a group chat, each told to act human and figure out who's an AI. After five rounds, each one privately votes on the others. It's a fun, low-cost way to see how models talk, bluff, and catch each other.
Source: DEV ·
On this virtual altar you add a photo and small memories of someone you've lost, like their favourite tea or song, light a diya, and an open model writes a short letter in warm Hinglish. It uses only the details you typed and never pretends to be the person.
Source: DEV ·
Upload a Markdown file and this tool checks every link, looks up old copies of broken ones in the Internet Archive, and has AI compare what the page said then and now. It shows its evidence and never quietly swaps links for you, so the fix stays your call.
Source: DEV ·
In this free game, agents take turns drawing with simple SVG shapes and guessing what the others drew, with points for harder clues. You can connect your own agent through MCP without an account, or just watch a live game. Some drawings are impressive, others hilariously bad.
Source: Hacker News ·
Paste a GitHub repo and this tool reads its imports, models, and cited papers, then matches each piece to the research behind it with a one-line key idea. Claude and Codex both pick the papers and links, and solid lines mean they agreed. You can trace an idea back through earlier work and export it all as BibTeX.
Source: Hacker News ·
This local agent summarizes a folder of files with Gemma, but every step is saved as it goes, so if you kill it mid-sentence a fresh process picks up at the exact step where it died. No finished work gets redone, and no model calls get paid for twice. A dashboard even has a big button to kill it yourself and watch it come back.
Source: DEV ·
Tired of AI-written sales emails buzzing their phone, a founder worked with Gemini to write a free Google Apps Script. Once a minute it moves any inbox email from someone they've never emailed into a separate folder to review later. They chose not to have AI read their emails, keeping it simple and predictable.
Source: Hacker News ·
A link checker only asks if a page loads, so links whose domain expired and got bought by someone else still pass. This free GitHub Action has Gemma compare what each link shows today with the archived page from when it was added, then opens one pull request that swaps bad links for the archived original. Code refuses any replacement the model didn't actually open.
Source: DEV ·
A firm that works under strict NDAs needed faster proofreading for reports due late on Friday, but couldn't send them to a cloud AI. So they set up open models on their own machines, and a local model now checks drafts for mistakes while people still decide every change. The write-up covers the engines, models, and hardware they tried.
Source: Hacker News ·
A developer rebuilding an old project wanted to write the code himself, so he worked with ChatGPT to write a Claude Code skill that sets four levels of help, from explaining things to building a whole feature. He stays mostly in "explain" and "small piece" mode and keeps the rules out of his main CLAUDE.md, so they only apply when he wants them. The skill file is shared.
Source: Bluesky ·
Someone used Claude Code with MCP to build an internal app in under 30 minutes where changing a process step also updates its diagram, so teams can see what's approved and compare past versions.
Source: X ·
This local tool reads emails you export from your inbox, sorts them into applications, interviews, rejections, and offers with an open model on your own machine, and keeps a simple tracker of every company and role. It flags applications that went quiet for a week so you know who to nudge. Nothing goes to the cloud.
Source: DEV ·
DMARC reports tell you who is sending email using your domain, but they arrive as piles of XML. This open-source app checks and sums them up with plain code, then an optional local model writes a readable briefing. The facts come from the code, not the AI, and nothing gets uploaded.
Source: DEV ·
Record an elder telling a story, or drop in an old voice note, and this offline app writes it down with Whisper, translates it for grandkids with a local Gemma model, and tags the people, places, and foods. Ask something like "what did she cook for Diwali?" and it answers with citations you can play back at the exact second she said it. It never pretends to be her.
Source: DEV ·
A developer hid a prompt injection inside a fake production log and asked different models to analyse it. Several free open models dropped the analysis and followed the hidden instruction, while the Gemini and Codex models tested did the real job. A good reminder to treat logs and other data you paste into AI as untrusted.
Source: DEV ·
Point this small tool at your own EPUB library, pick an author like Poe, and ask a question. A local Gemma model answers using only passages from that author's books, with the book and chapter cited under each answer. If the books don't cover it, it simply says nothing instead of making something up.
Source: DEV ·
Point this command-line tool at a line nobody understands, and it digs through git history, including times the code was deleted and brought back. A small local model then explains it in the voice of the old commit messages, and every commit it mentions is checked against the real history so you can tell facts from guesses.
Source: DEV ·
Type a place and a local Gemma model searches the web, picks a legend, and retells it as a short candlelit tale with sound effects. Each story ends with its sources and a label saying whether it's documented history, local legend, or internet lore, so you know what's real.
Source: DEV ·
This offline project reads one night of room sensor logs, works out which rooms were likely occupied minute by minute, and flags anything odd, like activity in a room nobody walked to. A local Gemma model then retells the night as a short ghost story, and a fact check makes sure it only quotes real times and readings. The demo uses made-up data.
Source: DEV ·
Tap one of seven traditional offerings on this illustrated altar to read its meaning, history, and how families do it differently, then ask your own question in English or Spanish. Gemma answers from a curated knowledge base, and a labelled guide takes over if the model isn't available.
Source: DEV ·
In this small Day of the Dead game you play a skeleton mail carrier helping lost ghosts cross over. Every ghost's backstory, dialogue, and letters were written by Gemma, with a batch pre-made so it plays instantly and an option to generate new ghosts live.
Source: DEV ·
This open-source caption tool runs audio through up to three speech-to-text services, and when they disagree it re-checks the short clip and has an AI decide what was actually said. It then cuts captions to fit any screen shape, keeping names and numbers on one line. It costs more per video, but means far less manual fixing.
Source: Hacker News ·
This free browser planner lets you sketch furniture or upload a photo of a piece you like. An AI turns the photo into an editable model in millimetres, and the app gives you a numbered list of boards to cut. Double-check measurements before you head to the saw.
Source: Hacker News ·
Someone who gets clients from helpful conversations online had Grok Bot build a daily morning digest of community and social threads they could add something useful to. It also works as ongoing market research.
Source: X ·
An artist used Claude Code with Maya's Python scripting to build a face rig for a cat knight character, with blinking eyelids, jaw and mouth movement, twitching ears, and a control board. Lip shapes are next.
Source: X ·
Wrap any command with this tool and, if it crashes, it gathers the error, traceback, recent git changes, and versions, then a local Gemma model writes a short report with the likely cause and a fix to try. The raw evidence is saved under the story, so you can check where the small model got it wrong.
Source: DEV ·
After a free bundle quietly rolled into a full-price renewal, someone built a small free tool where you list your subscriptions and how much you use them. Open models running on your own machine flag plans that are too big for you, ones you stopped using, prices that crept up, and services that overlap, and show what each one costs you a year. It never asks for your bank login, and the final call stays yours.
Source: DEV ·
For tired parents whose kids want one more story, this free app reads a Bengali story aloud sentence by sentence, with natural pauses. It uses a small open text-to-speech model that runs on a normal laptop with no internet, no GPU, and no subscription.
Source: DEV ·
In this free browser game you start by answering tasks yourself, save up to rent a GPU, and then hire a real small Qwen model running in your browser. Upgrades like structured output or date tools actually change how the model behaves, and you can write your own prompts.
Source: Hacker News ·
This open-source library has an AI interview your analysts one area at a time, informed by real query logs, and writes down what they know in short markdown and YAML files kept in git. It builds test questions as it goes, so you can check whether coding agents like Claude or Codex answer business questions better with the context.
Source: Hacker News ·
This open-source app checks in with you once a week. If you stop answering, it asks two people you trust, and only when both confirm does it release a personal letter to each and let them ask your notes questions like where a policy document is, with an open model citing the exact note. Plain code, not the AI, decides who sees what. It doesn't replace a proper will.
Source: DEV ·
A developer noticed coding agents always say the bug is fixed, so they built a free tool that turns past bug fixes in your repo, the ones that came with tests, into a test set. On their first real run, 27 of 66 patches the agents called done were wrong. It's a way to measure agents on your own code instead of public benchmarks.
Source: DEV ·
A developer set up a lead coding agent plus helpers for design, code review, SEO, and docs in OpenCode, with a rule that every change is proposed first and approved by them. The team built a complete gym website template from an empty folder, and it's now listed among the official Astro themes.
Source: DEV ·
Someone had Grok Bot pull highlights from their X posts, design a coffee table book around them, and place the print order, all in under a day. The finished book has now arrived at their door.
Source: X ·
Someone moved their personal agent off the cloud and onto a home storage box using the open-source OpenClaw harness. Its rules, personality, and memory live in plain text files kept in a private git repo, and every Monday morning it pulls their Notion tasks into a tidy email recap, with a check so it never sends the same one twice.
Source: DEV ·
This tool sends AI personas through your site in a real browser set up like their device, such as a slow phone, a screen reader, or high zoom, each with a goal like placing an order. A local Gemma model writes what went wrong in their words, but a complaint only counts if the browser measured it, like load time, tiny tap targets, unlabeled buttons, or low contrast.
Source: DEV ·
This video walks through a free little web app built inside Google Sheets. You take a photo of a paper receipt, Gemini reads the store, date, total, and category, and a new row appears in your budget. Gemini writes all the code from one message, and it runs on the free API tier with no server.
Source: YouTube ·
Give it a recorded family story and Whisper plus Gemma, both running locally, return a timestamped transcript, the people and places mentioned, a short summary, quotes checked against the recording, and a few questions to ask next time. The recording never leaves your machine.
Source: DEV ·
Type in the address of a defunct site and this app pulls its snapshots from the Internet Archive, then a local Gemma model tells the site's story and answers questions. Every sentence links to the archived page it came from, and if the archive doesn't hold the answer, it says it doesn't remember.
Source: DEV ·
Jot a few lines about a person, pet, place, or family ritual, add a photo if you like, and a local Gemma model shapes it into an editable keepsake card with a short reflection and a question to come back to. It's told not to invent dates, names, or quotes, and everything stays in your browser with an export option.
Source: DEV ·
A parent built a Grok bot that takes a photo of their 6th grader's study guide plus the grade level and turns the lesson into a cartoon that follows the guide in order, ending with a recap of key terms.
Source: X ·
Someone connected Claude to Blender and built a 3D model of a whole zoo, with paths, buildings, and enclosures, that you can explore by walking in from the entrance on your laptop.
Source: X ·
Running several coding agents at once, someone kept losing track of which one needed them. Their free, open-source app shows sessions from tools like Claude Code, Codex, and Cursor in the Mac notch, so you can answer prompts, see costs and limits, and send back failed builds. Data stays on your Mac.
Source: X ·
This free, open-source app downloads a local AI model and a library of emergency guides, repair manuals, and Wikipedia so you can ask questions like how to keep food safe in a blackout with no internet. Every answer shows the articles it used so you can check them. Set it up while you're online, and in a real emergency always follow official guidance and contact emergency services.
Source: Hacker News ·
Needing a 15-second ad for a side project, someone had Claude Code write a small renderer that reuses the app's own design. They described changes in plain language and re-rendered, never opening a video editor, and ended up with translated versions and even tests that check the timing of each scene.
Source: Hacker News ·
Using Codex, someone built a low-poly 3D spot where you pick a wall, choose stencils, adjust the spray, add paint drips, and even shake the can. The favorite touch is invisible paint you reveal with a UV flashlight. A fun, legal way to play with street art.
Source: X ·
Point this free tool at a SQL migration and fixed rules flag things that can hurt in production, like dropping a column, adding a required column to a huge table, or building an index that blocks writes. A local model can then explain what might break, who it affects, and a safer rollout order. The rule scan stays the source of truth.
Source: DEV ·
This command-line tool reads a script as text, never running it, and flags risky patterns with a severity level and the matching attack technique. A small model running locally through Ollama then explains the findings in plain words and can export a short report. It's a learning and triage aid, not a replacement for proper security tools.
Source: DEV ·
A busy student is building a desk gadget where you describe all your commitments in one go, review the organised plan a small local Gemma model suggests, and get reminders on a little LED screen as deadlines come close. It's an early project, but the idea is to spend less effort managing tasks and more doing them.
Source: DEV ·
Someone who can't play the piano wanted to hear sheet music played with real feeling. They used Claude to turn music braille, which already carries tempo, dynamics, and expression marks, into a MIDI performance of a 19th-century piano piece, and shared the result as open source.
Source: Hacker News ·
A developer who doesn't write Ruby used a coding assistant to fix a small bug in Forem, the open-source platform behind dev.to. The code was the quick part; most of the 12 weeks went into waiting, review, and learning the project's rules. A realistic look at contributing to open source with AI help.
Source: DEV ·
During class, this companion transcribes the lecture and keeps linking what's being said to related moments in past lectures and slides. Claude explains or summarizes on tap, Gemini searches old lectures by meaning, and after class it quizzes you out loud on what was just covered. Check your school's rules on recording first.
Source: DEV ·
Paste the message and Gemma drafts a reply of 60 words or less, noting the one fact it must not make up. A local check then blocks the draft if it adds any number, email, or name that wasn't in the original message or your note.
Source: DEV ·
This small Mac menu bar app checks every screenshot you take for things like API keys, passwords, emails, phone numbers, and card numbers, and puts a blurred copy on your clipboard. Simple rules catch the obvious stuff and a Gemma model running locally decides the unclear cases, so nothing leaves your Mac. You can click the preview to fix anything it missed.
Source: DEV ·
Type something like "I have an exam tomorrow, find time to study" and this open-source planner reads your Google Calendar, finds a realistic gap, and proposes a focus block with a reason. Nothing is written to your calendar until you approve it. It can run on a local Gemma model, even on a small Arduino board with status lights.
Source: DEV ·
This open-source food diary app lets you press the iPhone Action Button and say something like "log my breakfast", and its voice assistant adds the meal for you. By default it shows a confirmation card before saving, with undo afterwards. Voice mode needs your own AI API key or the app's paid plan. It's a logging tool, not nutrition advice.
Source: DEV ·
Paste a notice, snap a photo of a printed one, or give a link, and Gemma pulls out each action, who it's for, the deadline, and any fees or documents mentioned. A separate check makes sure every item is backed by the notice's own words and flags anything missing for a person to review.
Source: DEV ·
With one prompt, a developer had Claude build an interactive 3D workbench where a card rides a little shuttle through a gate for login checks, drawers for the database, and a cache that keeps its copy. They then refined it in a few rounds of back and forth. A playful way to explain a technical idea instead of drawing boxes and arrows.
Source: DEV ·
This local app reads the text in a folder of screenshots so you can search it, and a Gemma model can suggest categories you choose, each with a quote from the screenshot as proof. You approve every change, files are copied rather than moved by default, and it all runs on your own computer. It's slow, but private.
Source: DEV ·
Record a family member talking while they cook, and Whisper plus a local Gemma model turn it into a proper recipe with amounts, steps, tips, and the stories they told along the way. When something is vague, like "a spoon of" something, it writes down a question to ask them next time instead of guessing. Everything runs on your own computer.
Source: DEV ·
Searching for a good first issue usually turns up ones that are closed, assigned, or already claimed in the comments. This free tool collects beginner issues from a set of AI projects, re-checks them right before showing you, drops any that are taken, and has a local Gemma model explain the top five with a difficulty rating and first steps.
Source: DEV ·
This Chrome and Edge extension finds duplicate bookmarks, suggests a handful of folders that fit what you've actually saved, and sorts everything into them with Gemma running inside your browser. You can drag things around before applying, it saves a backup so you can undo everything, and new bookmarks get filed as you save them.
Source: DEV ·
After a free trial quietly turned into a plan with an early exit fee, a developer built a Chrome extension that reads the terms around a sign-up or checkout button. A Gemma model running in the browser pins a card listing things like exit fees, lock-in periods, and trials that turn paid, each linked to the exact sentence. If you sign up anyway, it adds the cancel-by dates to your calendar. It's a reading aid, not legal advice.
Source: DEV ·
Snap a photo of the envelope or letter and drop it in a folder. Overnight, a local model reads it and decides whether it needs action. Most of the time it files it quietly, and in the morning you get a short note only for the ones that matter, with the deadline and a slow read-aloud. Nothing leaves your computer.
Source: DEV ·
Paste in your bug and this free command-line tool won't fix it. It uses Claude, or a local model, to question your assumptions and point you at the evidence you've been skipping, so you find the fix yourself and understand your own code better.
Source: Hacker News ·
Someone connected Claude Code to the budget Android head unit in their car and, by their account, found an always-listening assistant eating CPU, preinstalled logging apps, and a remote updater. They had it back up and remove the unwanted apps, block trackers, update old components, and even recalibrate the backup camera lines. Only try this on hardware you own, and back everything up first.
Source: X ·
A developer built a free, open-source connector that lets ChatGPT or Claude search a food delivery app, fill your cart, check out, and track the delivery, all in a conversation. Since it places real orders, double-check the cart and total before you confirm.
Source: X ·
A database quietly crashed and nobody noticed for five days, because no one reads hundreds of routine log lines every morning. Now a local model reads the logs overnight and points out only the lines that need a person.
Source: DEV ·
Coming back to a project means piecing together what you were doing. This app reads your git history, open changes, pull requests, and past coding agent chats, then a local Gemma model writes a briefing that starts with your next step and links each point to its source.
Source: DEV ·
Type something like "remind me to drink water" or "track what I spend" and this Android app builds a small working tool for it, like a counter, reminder, timer, or bill splitter, that you can preview, save, and reuse. It runs Gemma on the phone itself, so after a one-time model download it works offline.
Source: DEV ·
Keeping up with F1 meant five tabs and a lot of cross-checking. This dashboard collects race news, groups articles about the same story, and writes a briefing with Gemma. You can also ask questions, and answers separate what was reported from what is only a guess, with links to the sources.
Source: DEV ·
Paste a Flutter or Dart error and a small Qwen model running on your machine suggests a likely cause and fix. Once you've checked a fix and marked it as verified, the tool remembers it and brings it back the next time the same error shows up.
Source: DEV ·
This free, open-source desktop app works with your files much like the Claude and ChatGPT desktop apps do, but the model runs on your own machine. It picks a Gemma or gpt-oss model that suits your hardware, reads and edits files in a folder, looks at images, and writes small sandboxed scripts for things like converting documents or drawing charts. It also works offline.
Source: Hacker News ·
Most saved videos never get rewatched. Each night this tool takes short tech videos from your saved posts, has Whisper write out what's said, and has a local Gemma model turn it into a short note with the tools, numbers, and a quote as proof. If a video has no speech, Gemma reads a few frames of the screen instead.
Source: DEV ·
A favorite pizzeria had no app, only a phone line. Say what you want and a spending limit, and this assistant uses Gemma to understand the order, calls the restaurant with a real voice, says up front that it's an AI, checks the price stays under your limit, and only shares your address when asked.
Source: DEV ·
Re-running a failed test throws away the evidence. This open-source tool keeps every CI test report, rebuilds each test's history of fails and passes, and lets fixed rules decide if it's flaky. A local Gemma model can add a clearly labeled guess at the cause, but it never makes the call.
Source: DEV ·
Before closing the laptop, this desktop helper looks at what you meant to do, what changed in your code, and which checks actually passed. A local Gemma model writes a short closing card about what's done, what's still uncertain, and where to start next time, without committing or pushing anything.
Source: DEV ·
A job seeker kept losing mornings and sleep to checking email "just in case". They built a small board on their desk that reads their inbox overnight, lets Gemma decide on the device whether anything needs them, and shows one simple face. In the morning it also counts down to their bus.
Source: DEV ·
Describe what you feel like eating, your budget, and any diet preferences, and this tool searches for dishes, filters out ones that don't fit, and ranks the rest with some variety. A local Gemma model suggests extra search ideas, and uncertain prices or nutrition estimates are clearly marked as such.
Source: DEV ·
Drop exports from your bank, smart meter, or fitness tracker into a folder. Overnight, a local agent forecasts the next month, looks for odd readings like a possible water leak, and checks for price hikes, forgotten subscriptions, or double charges. On a quiet morning it says almost nothing.
Source: DEV ·
Paste the assignment instructions and an open model turns them into a checklist. Upload your files and the app flags things like a report over the page limit or a missing screenshot, leaves judgment calls marked for you to review, and packs everything into a tidy ZIP.
Source: DEV ·
Snap a photo of each page you've marked with a highlighter. Overnight, the app finds the highlighted passages and Gemma types them out, so in the morning you just keep, edit, or delete each quote. The page photos are deleted after the run.
Source: DEV ·
A cyclist kept going over their easy heart rate zone and wanted to know why. Drop in a ride file and this tool, built with the open TabPFN model, splits the extra minutes between the steep road, the heat, and the rider. Give it a route and weather forecast and it shows how much can be ridden easy. It gives no training advice.
Source: DEV ·
One small edit and a Word file grows random gaps, page breaks, or a paragraph that won't line up. Upload the document, point at the problem, and describe it in plain words. A local Gemma model works out why it happened and suggests a targeted fix, and the file stays on your laptop.
Source: DEV ·
Instead of sending every messy expense export through a model, a local Gemma model looks only at the column names once and writes a cleanup recipe. After checks and your approval, plain code runs that recipe every month, and it refuses files whose layout has changed instead of quietly getting them wrong.
Source: DEV ·
A student lost marks for AI notes that were correct but not what the teacher covered. This study workspace uses Whisper and Gemma to turn lecture recordings and class PDFs into notes based on that material, makes flashcards, helps with lab records, and practices the questions an examiner might ask about your own work.
Source: DEV ·
The reasoning behind choices made with an AI usually vanishes when the chat ends. This free framework has the AI write short decision records in your git repo listing the options, its recommendation, how confident it is, and what would change its mind, so you approve or override each one and the reasons stick around.
Source: Hacker News ·
While results were coming in for a provincial election, someone used Claude Code to build an interactive dashboard that pulls the latest numbers from the official results page as they update. It's a quick way to follow a big night with clearer charts than the official tables.
Source: X ·
Two co-founders kept losing track of things in their WhatsApp group, and the good task apps were paid. In a day they built their own with Claude Code, with Claude sitting in the chat to turn messages into tasks and send reminders.
Source: X ·
Instead of opening every class tab, this planner pulls due dates and links from your classes and has a local Gemma model suggest an order and study session lengths. The real due dates always stay visible, and if it can't tell whether something was handed in, it says so.
Source: DEV ·
This browser extension hides distractions on YouTube so you can focus on learning. When you do want to know what viewers think, a small Gemma model running on your computer sorts up to 100 comments by tone and gives you a short summary of the reactions.
Source: DEV ·
Paste a messy trip chat and a local Gemma model only picks out the expenses, who paid, and who should be left out. Plain code then does the actual math, keeps uncertain items like "maybe 350?" separate, and checks that every balance adds up, since chatbots often slip on exactly this.
Source: DEV ·
Upload a class recording in Hindi, English, or a mix of both and get a timestamped transcript, a summary, likely exam topics linked to where they were taught, flashcards, and any deadlines the teacher mentioned. Whisper and Gemma run on your own computer by default.
Source: DEV ·
A student preparing for entrance exams had already watched one session and wanted to know which parts of another video were actually new. Paste two YouTube links and this tool shows what looks familiar, what seems new, and the evidence behind each call, so you only watch the parts that add something.
Source: DEV ·
When agents write and review code, the final approval still falls to a person, and it easily turns into a skim. This free open-source tool runs after your AI review and tells the approver how risky the merge is and which files carry that risk, so attention goes where it matters.
Source: Hacker News ·
Reading a posting properly took one developer about 17 minutes, mostly to answer three questions: is it truly remote, what hours, and how much experience. Their tool has a local Gemma model suggest those answers and then plain code double-checks them against the posting text.
Source: DEV ·
One wrong digit in a deed can leave a lot boundary that doesn't close, and nobody notices until the surveyors show up. A developer fine-tuned a small Qwen model for about a dollar to read deed lines, then plain code checks the shape closes and flags the exact line for a person to review instead of guessing.
Source: DEV ·
Upload a CSV or Excel file and a local Gemma model splits the boring first look into jobs: spotting outliers and duplicates, drawing charts, explaining findings in plain English, writing the Python and R code, and suggesting clean-up steps. It all runs on your machine with no account, wrapped in a playful pixel-art office.
Source: DEV ·
Instead of answering every email faster, this agent guesses which ones would cost you something if ignored. The top few get a two-line summary from a local Gemma model, the rest are counted, and the least important are parked for a week and only come back if the sender follows up.
Source: DEV ·
"I'll send it tomorrow" is easy to type and easy to forget. This tool reads an exported group chat on your own laptop, picks out the promises you made, works out which are late, and drafts a quick catch-up reply. A local Gemma model does the reading, so the chat never leaves your computer.
Source: DEV ·
Tired of an inbox mostly full of CI noise, a developer built a laptop tool that reads GitHub notifications overnight with a local Gemma model and decides which ones need them. In the morning an open voice model reads a two-minute summary, with a page of links for anything worth opening.
Source: DEV ·
Type a short brief before bed and a local Gemma model drafts five post ideas with titles, outlines, and suggested posting times. A second model trained on real posts scores each one, and in the morning you're told which ideas are strong, or that none are and you can go back to sleep.
Source: DEV ·
In Pakistan, going even one unit over a monthly limit can raise the next six bills a lot. Read your meter once and this open-source forecaster tells you how many units a day you can use to stay under, staying quiet unless the risk is real. A local Gemma model can also read the meter photo and bill for you.
Source: DEV ·
Stage your changes, run one command, and a small model built right into the tool suggests a commit message in the same format your project already uses. Nothing is committed without your keypress, and it can also draft a changelog and version tag for a release. Your code never leaves your machine.
Source: DEV ·
Most open tabs stay open for one small detail, like a flight time or a comparison you meant to finish. This tool predicts which tabs you'll never read, sets those aside only when it's safe, and gives you one calm page of takeaways plus a short spoken summary, using open models like Gemma.
Source: DEV ·
Somewhere in hundreds of flatmate and family messages is the one that needs you, like being home for the internet technician. This tool reads exported chats with a small local Gemma model each morning and gives two lines per chat: what happened, and what you need to do, or "nothing for you".
Source: DEV ·
Run one command and a small local Qwen model reads your recent commits and writes a ready-to-paste update split into yesterday, today, and blockers. It saves the morning scramble of scrolling through history to remember what you did.
Source: DEV ·
A creator used Claude Code to build a one-on-one wizard battle that runs in the browser with WebGPU. Spells fly as beams of light, attacks can clash, arenas break apart, and you can play with a real console controller.
Source: X ·
A local tool reads sales, stock, purchase order, and invoice spreadsheets and flags the few rows that need attention, like stock running out before the next delivery or duplicate invoices. A local model explains each flag and suggests a next step, and the owner reviews everything before acting.
Source: DEV ·
Nobody reads a 9,000-word terms of service, so a developer fine-tuned a small Qwen model on a public dataset of unfair contract terms. Paste in the text and it highlights clauses worth a closer look, like forced arbitration, one-sided changes, or limits on liability. It's a reading aid, not legal advice.
Source: DEV ·
A tech reviewer used Claude Code to remove unwanted apps, turn off ads and tracking, stop needless background processes, and clear temporary files, all by describing what they wanted instead of hunting for commands. If you try it, make a restore point first and review each change before saying yes.
Source: Blog ·
Lots of people dump links, shopping lists, and reminders into a chat with themselves and never look back. This browser tool reads an exported copy with a local Gemma model and gives you one card: what's due, a merged shopping list, what's safe to forget, and sensitive bits like passwords and codes you should delete.
Source: DEV ·
Say a wake word and tell your laptop what you want in your own everyday way, even in Nigerian Pidgin, like "go back small" or "make am louder". Whisper hears it, a small Gemma model turns it into one command, and the video rewinds, skips, or gets louder. It all runs offline with no account.
Source: DEV ·
Faced with a late-night pile of two dozen support tickets, a developer built a tool that sorts them by urgency, matches each to the right playbook, adds labels, and drafts the first replies. It can run on a local Gemma model, and nothing goes out until a person approves it.
Source: DEV ·
After a rough month of poisoned packages, a developer built a tool that reads every package, agent skill, MCP server, and hook file in a project overnight without running any of them. A local model asks one question of each: does this do what it claims? You wake up to a list of anything suspicious.
Source: DEV ·
An artist couldn't get a lamp and its light right in their painting, so they asked Grok to adjust that part of the image. They used the result as a study guide to finish the real painting by hand.
Source: X ·
Paste your notes and a small local model writes the flashcards, a spaced-repetition schedule decides what to review, and you answer by voice or typing while the model grades you. It runs fully offline.
Source: DEV ·
Unhappy with the voices and clutter in other read-aloud extensions, a developer vibe-coded an open-source one that reads web pages and the online Kindle reader. The voices come from open models that run right in the browser, so there's no sign-up.
Source: Hacker News ·
Tap the mic on your phone, say what's on your mind, and Whisper turns it into text in the browser. Later a small Gemma model on your own computer gives it a title, a category, tags, and a due date if you mentioned one, while keeping your original words. You can also ask it questions like who you need to call.
Source: DEV ·
When a favorite channel posted dozens of talks in a month, a learner built a tool that turns each new upload into a short card with key claims and exact quotes, then merges repeated ideas into one map. Each morning a local model adds only what's new to their own notes, and quotes are kept only if they match the transcript word for word.
Source: DEV ·
A longtime flashcard user built an app that reads their deck's questions through headphones, listens to spoken answers, and grades them, so reviews can happen away from a screen. Screen reviews are free; the voice tutor is a paid subscription after a trial.
Source: Hacker News ·
One person asked Grok to design a cabin for them using only their own posts as a guide. It came back with a home of natural materials and a big rose garden, a fun way to turn your interests into a design idea.
Source: X ·
Paste a week of volunteer chat messages and a local Gemma model picks out the phrases that matter, like "Tuesday" or "after 3", while plain code turns them into real dates. You accept or reject each suggestion next to the message it came from, so changed plans never slip through.
Source: DEV ·
A student who kept missing class updates built a local tool that checks Google Classroom and email, throws out the noise, and sorts real deadlines into panic, coming up, and relaxed buckets. A small Gemma model does the reading, and you get a one-minute spoken summary instead of a crowded inbox.
Source: DEV ·
During a week full of "patch now" posts, a developer built a laptop tool that scans their projects against the latest advisories overnight and leaves a two-line summary and a short checklist. Plain code decides what's true and a local Gemma model only writes the wording, with a checker that rejects any number or version the model made up.
Source: DEV ·
After a fix, an open-source tool reads the diff, the new test, and the commit message, and Gemma writes up the symptom, root cause, and lesson. Before your next commit it checks whether the change repeats a past mistake and points to the lines, so lessons don't stay buried in git history.
Source: DEV ·
An iPhone app reads your screenshots with a small vision model that runs on the phone itself and turns each one into a card: an event for your calendar, a place for maps, or a receipt for an expense list. You just swipe to accept or skip, and it also gathers blurry or duplicate shots for a quick clean-up.
Source: DEV ·
Take up to three photos of your fridge and pantry, and a vision model checks what's full, low, or empty. A local text model turns that into a sorted shopping list plus a ready-to-send message for whoever's doing the shopping, and staples you always buy get added automatically.
Source: DEV ·
After waiting over a year for one course to be recognized at a new university, a student built an open-source helper that compares two syllabi topic by topic with a local Gemma model. It marks each topic as covered, partial, or missing and drafts the report, while the final decision stays with a person.
Source: DEV ·
A developer built a terminal app for running several Codex and Claude Code sessions side by side. It shows what each agent is doing in a live feed, jumps you to the next one that needs attention, and uses a cheap model to label links, branches, and sessions for you.
Source: Hacker News ·
Someone had Claude review a water-flow simulation they wrote in C as a math intern back in 2013. The result is a cleaned-up open-source solver with docs on the theory, a getting-started guide, and validation tests that compare it against known results.
Source: Hacker News ·
A beginner-friendly walkthrough shows how to teach Claude Code to generate images, voices, and music and stitch them together with FFmpeg. It ends with a 101-second narrated story made for well under a dollar in generation costs, and every prompt is shared.
Source: DEV ·
A developer built a small pipeline that reads new posts from public job forums, keeps only the ones where someone is actually hiring, and adds a short summary, the pay as numbers, and a scam score. The post shares the prompt rules that made the biggest difference.
Source: DEV ·
A local command-line helper records failed commands and the fixes you save. When an error shows up again, it tells you what worked before, and for brand-new errors a local model explains the failure in plain words. Nothing leaves your machine.
Source: DEV ·
A data science researcher tried a Claude Code plugin for updating legacy software. It maps the old code, writes testable rules with citations, waits for your approval, then rewrites and compares old and new versions on the same inputs, which is useful for lab code nobody maintains anymore.
Source: Substack ·
One user had Claude walk them through installing Linux as a dual boot on a laptop, then troubleshoot a Wi-Fi card problem and a Windows sandbox. They noted it was much weaker at sorting things into fuzzy categories, so it's worth checking its work on tasks like that.
Source: Bluesky ·
You list dreaded tasks in plain words, and a small local model breaks each one into steps marked for the machine or for you. It then drafts the reply you've been avoiding, sorts a messy folder, or pulls action items out of meeting notes, all offline.
Source: DEV ·
A developer revived an old hobby question: can a program beat a million points in the puzzle game Drop7? The open repo is set up so Codex or Claude Code can run experiments, audit results, and log evidence, and people can also submit their own human strategies to a leaderboard.
Source: Hacker News ·
Because agents now fill in links, quotes, and examples in their docs, a writer built a small Mac markdown editor that merges an agent's changes into the open file in real time. Your unsaved work stays put, and undo only touches your own edits.
Source: Hacker News ·
A developer built a Windows helper where a pixel-art hero pops up to remind you to drink water, rest your eyes, or stretch, with a fresh line each time in the tone you pick. It runs on a small local model, is set up with one plain sentence, and holds reminders while you're in a video call.
Source: DEV ·
Paste a repo link or a readme, answer a few questions, and a group of agents suggests names in different styles, checks which domains are free, and lets you tweak the list by chatting. It's open source and runs on Gemma.
Source: DEV ·
A developer built an on-device tool that reads incoming notifications and emails, scores how urgent each one is with a small local model, and shows a short summary every few hours. Only items that pass a strict emergency bar break through, and it keeps a log of what it held back.
Source: DEV ·
Facing over a thousand saved articles, a developer built an offline agent that learns from what they actually read, keeps a handful worth reading, and clears the rest. A local model then writes a short digest of the keepers, with no account or cloud needed.
Source: DEV ·
Instead of letting AI silently rewrite a draft, an open-source Google Docs add-on uses Gemini to leave suggested edits and comments, each with a reason and a short learning point. It can also review tables and charts, so you improve as a writer rather than just copying fixes.
Source: DEV ·
An automation freelancer whose clients explain requests in long voice messages built a laptop tool that transcribes them with Whisper and has a local Gemma model pull out the trigger, steps, and missing details. Regular code then builds an importable n8n workflow plus a list of follow-up questions for the client.
Source: DEV ·
A teacher of AI topics built a web app in Google AI Studio that takes a spoken or typed idea, has Gemini write a four-panel storyboard with dialogue, and renders matching panels in a consistent style. Handy for explaining technical ideas in classes and blog posts.
Source: DEV ·
A game developer used Codex and other AI tools to rebuild their own 2012 iPad game from the ground up as a full PC release in about three months. The remastered version is launching on Steam.
Source: X ·
A team that ships several apps built a Telegram bot with Claude that gathers overnight data from their app store, code, task, crash, and support tools. Each morning it posts one short summary, so they don't have to open five dashboards to see what happened.
Source: X ·
A free tool writes a handoff note from your Claude Code, Codex, or Gemini CLI chat, with the task, recent requests, changed files, and test command, and hides secrets first. Another assistant can then pick up where the first one stopped.
Source: DEV ·
After making a pile of architecture decisions in a few days, a developer built a small tool so that saying "record this" in Claude Code writes the decision, and the options they turned down, into the repo as a markdown file, so the reasoning isn't lost later.
Source: X ·
A maths teacher made an integration quiz with Gemini and used it with older students right after their first lesson on the topic. They found it built confidence and showed who had understood.
Source: Bluesky ·
After a coding agent helped fine-tune Japanese typesetting and accessibility on a personal site, the builder had it break those details into separate rules, research how experts judge each one, and add them as tests. That way the polish doesn't quietly break with the next style change.
Source: DEV ·
A web developer set up one main agent that builds features and four helpers it calls on for design review, code review, SEO checks, and keeping docs up to date. Each helper is a single markdown file, and nothing gets built until they approve the plan.
Source: DEV ·
An open-source app reads the PDFs and receipts you drop in, pulls out details like issuer, amount, and due date, and suggests clean file names and folders. You review everything side by side before files move, it can undo, and it all runs on your own computer.
Source: DEV ·
Tired of practice apps swapping original songs for covers, a hobby singer had Claude Code help build a Mac app that listens to a track once and then scores their pitch and timing as they sing along. It doesn't save the song, and it's open source.
Source: Hacker News ·
A free command-line tool scans a Python project, draws how its files depend on each other, and flags problems like circular imports. A small open model running locally then explains the architecture in plain language.
Source: DEV ·
The maker of a vocabulary app shares the checks they use on AI-written example sentences, after a nonsense word produced a card made entirely of the model's apology. The post covers invented words, refusals that slip through, and what the checks still miss.
Source: DEV ·
A developer wrote a free add-on for Claude Code that checks each request before it's sent. It suggests compacting when the context is huge and pauses new helper agents when you're near your limit, and over 75 days it cut their biggest bursts sharply.
Source: Hacker News ·
A team built a physics simulation of a thruster-powered car doing stunts in gusty wind and tested how well language models could fly it. Big models were too slow to react, while a small model they trained with reinforcement learning landed far more often.
Source: Blog ·
A free agent skill has your coding agent explore your app as an impatient skimmer, a skeptic who checks pricing and privacy, a phone user, and someone who tries to break things. It reports where each one gets stuck before reaching the first useful result.
Source: DEV ·
A creator wanted a relaxing digital slime to play with, so they built a small slime simulator with Claude Code and shared a short clip of it squishing and stretching.
Source: X ·
A creator used Claude Code to write a program that generates every sound from math and draws all of the video's 11,000-plus frames in sync with the beat, with no music or video software involved.
Source: X ·
Out of nostalgia, a developer worked with a team of AI agents for about a week to build a Windows 95-style desktop, with movable windows, icons, and little apps, drawn entirely by the terminal and reachable over SSH. It's open source.
Source: DEV ·
When a severe PHP vulnerability started being exploited, the Internet Archive's operations team used Claude to write fixes for their older PHP versions and local models to run simulated attacks before deploying. They were back online on updated versions after a short outage.
Source: Bluesky ·
A developer couldn't remember how they'd gotten WebGPU working on Linux, so they sent an AI agent to search their own computer. It turned up the launch script and notes they'd written and backed up, and the post shares the setup for running AI models in the browser.
Source: DEV ·
A blogger moved a small job that runs a language model over their own posts from their laptop to a hosted service, keeping the same open-weight model, prompts, and code. Their notes cover setup, speed, and cost, including how prepaid credit simply stops calls when it runs out.
Source: DEV ·
A developer running several Claude Code sessions at once built a free terminal sidebar that shows each one as running, waiting, or done. Waiting sessions pulse and send a notification, and you can jump straight to the one that needs you.
Source: DEV ·
An open-source tool gives agents like Claude Code, Codex, or Gemini a sandboxed copy of your folder, tracks what every command really changed, and lets nothing reach the real files until you or a rule approves it. Every change can be undone.
Source: DEV ·
A free dashboard reads your local Claude Code and Codex sessions and shows this month's spend, how much of your usage limit is left, and what each session cost. It makes it easy to spot the runs that went in circles.
Source: DEV ·
A team asks its coding agents to record only the important things in a shared wiki as they work: decisions and who made them, open questions with an owner, and where each area stands. People read the wiki to set direction instead of rereading long agent transcripts.
Source: Hacker News ·
A creator built a browser fighting game where your hands, tracked through the webcam, move the puppets. The hand tracking uses a free Google library, and the game's physics were coded with Claude's help.
Source: Hacker News ·
A creator built a small game where you sail a boat around an autumn lake and collect kanji floating over the things they mean, like mountain, temple, and lake. It was made with Claude for about six dollars, and anyone can play it.
Source: X ·
A security firm released two free plugins for coding agents: one reproduces a reported bug, writes a fix, and tests it, and the other hunts for regressions the change may have caused. The project's own rules stay in charge, and a person reviews the results.
Source: Blog ·
OpenAI and research partners split a mouse-brain gene study across thousands of AI agents, each checking one gene change against the published literature. In about half an hour they pulled out leads that experts rated as new and worth testing. These are research ideas for lab follow-up, not treatments.
Source: Blog ·
A developer ran Claude Code's review modes on 20 hard pull requests from a public benchmark. The multi-agent mode caught a couple more serious bugs but took far longer and cost much more per review, and the post suggests which setting makes sense for which team.
Source: Blog ·
A free, open-source desktop app imports your browser bookmarks, writes short summaries, and lets you search them by meaning. It works with a local model or any compatible AI service, and keeps your data on your own device.
Source: Hacker News ·
A developer built a real-time card battle game with live matches, rooms you can share with friends, and a leaderboard in about two days with Claude Code.
Source: Hacker News ·
A free, open-source tool records your sessions across many coding agents and turns repeated fixes and workflows into reusable notes after you approve them. Any agent you switch to next can pick them up, so you stop re-explaining your project.
Source: Hacker News ·
A developer wrote a small open-source script that reads their dev.to account and emails a weekly digest of top posts from followed tags plus popular ones outside their bubble. An open-weight model adds a one-line summary to each card, and you can preview it without sending anything.
Source: DEV ·
A free, open-source helper waits until your computer is idle, spots clutter like duplicates and old installers, and reads documents with a local model to suggest where they belong. Everything runs on your machine, and every change can be undone in one click.
Source: DEV ·
An open-source on-call helper collects overnight error alerts, removes the repeats, and uses an open model to write a short morning summary. It can also prepare a tested code fix that waits for an engineer to approve it.
Source: DEV ·
A reader gave Claude 20 years of their own reading history and asked it to project when their to-read pile will become longer than they could ever get through. It's a fun way to look at your habits and decide what to prioritize.
Source: Bluesky ·
A writer used a model running on their own computer to come up with an obscure Linux boot setting that tuned their new laptop's power management beyond what the settings menus offer. It worked on the first try, after past attempts took days of trial and error.
Source: Bluesky ·
An indie developer had Claude generate a fireball animation with an image model and wire it into their desert racing game with 3D particle effects. A playable demo of the explosions is linked in the thread.
Source: Bluesky ·
An open-source desktop tool combines simple file checks with a small model running on your own computer to find duplicates and suggest tidy folders. Nothing is sent to the cloud, and every move can be reversed.
Source: DEV ·
Uber's engineers explain how they built an agent that finds feature flags that are fully rolled out and removes the leftover code paths and tests across many files. Safety checks, validation, and human review points are built into every cleanup.
Source: Blog ·
A travel service added a connector that lets you ask Claude for flights in plain language, compare options, and book one without filling in forms. The first booking goes through a secure checkout page, so you stay in control of payment.
Source: Hacker News ·
An open-source mobile testing tool now gives you a read-only page where you and teammates can watch an agent tap through your app on a simulator in real time. A running list shows every tap, swipe, and app launch as it happens.
Source: DEV ·
A free tool asks for a few rough notes when you stop working, adds a read-only snapshot of your code changes, and has a small local model turn it into a short summary with next steps. When you come back, you can pick up where you left off or paste it into a coding agent.
Source: DEV ·
A Windows specialist had Claude build, in one go, a desktop tool that reads antivirus performance logs and shows them as a timeline with charts, color-coded findings, and plain explanations. They said it captured years of their own troubleshooting ideas.
Source: Bluesky ·
A developer wrote down their open-source contribution process as rules and workflows a coding agent can follow, from finding a free issue to answering reviews. Nothing goes out without their OK, and in one week it helped them open 20 merge requests, most of them merged.
Source: DEV ·
Anthropic's security team had Claude map which parts of 25 open-source projects are reachable by default and flag code that didn't need to be exposed. It found hundreds of issues, and the post explains how developers can apply the same approach to their own projects.
Source: Blog ·
A community manager is building an open-source tool that saves good answers from Discord, GitHub, and Slack support chats so the team can review them and an agent can reuse them. It helps stop the same questions from being answered from scratch.
Source: DEV ·
A student built an open-source helper that looks at math you're writing by hand on a tablet, circles the first real mistake, ticks what's right, and gives hints instead of answers. It can also turn a step you box into a short animated explanation.
Source: Hacker News ·
A free, single-file tool runs Claude Code or Codex on a written brief while blocking it from everything outside the folders you choose. It stops the run once it hits the budget you set and logs what each run cost and whether it finished.
Source: Hacker News ·
A free starter repo shows how one team runs Claude Code in a sandbox with rules that auto-approve safe commands and linters that catch repeat mistakes. The team says the setup cut their token spend by around 80 percent.
Source: Hacker News ·
A developer built a browser-based, Command & Conquer style strategy game that works on phones, mostly through Claude Code over a dozen rounds of changes. It has 3D units, random maps, and even multiplayer between two nearby devices over sound.
Source: Blog ·
A free start page reads the blogs and feeds you follow and uses a small embedding model on your own computer to put the most relevant stories on top. It learns from what you open, like, and hide, and you can tune the ranking with sliders.
Source: Hacker News ·
A developer made a free, open-source quiz game where you race a vision model to answer questions about pictures, alone or with friends. The cards were made with help from vision models and Claude Code.
Source: Hacker News ·
A developer built a small app for a one-person wholesale bakery to count shelf stock at each cafe, build deliveries, send branded invoices, and track who owes what. An AI planning step interviewed them first and caught problems before any code was written.
Source: DEV ·
A free command-line tool reads the last day of commits in a repo and has a local model draft your standup notes. Every bullet has to point to a real commit, and any line that can't is thrown out, so made-up work doesn't slip in.
Source: DEV ·
Anthropic's research team, working with outside non-experts, had Claude design small proteins meant to latch onto scorpion toxins as an early step toward cheaper antivenoms. Labs tested 211 designs, 17 met the bar, and all the data was released openly. This is early lab research, not a treatment.
Source: Blog ·
A free Vite plugin lets you click any element in your running app, type the change you want, and hand your coding agent the exact file and line it lives on.
Source: Hacker News ·
A free terminal app lets you run multiple coding agents at once, each in its own copy of the repo, and switch between them from the keyboard. It works with whichever agent you already use.
Source: Hacker News ·
A creator uses a free Claude skill to turn one question into an interactive model of an object, like a Rubik's Cube or a pencil sharpener. You can pull it apart and run it step by step to see how the parts work together.
Source: X ·
Someone used Claude Code to work out how their own robot vacuum talked to its app after the maker's cloud went offline, then got it running again. It's a good example of rescuing hardware you already own instead of throwing it out.
Source: X ·
A developer built a kart racing game in Flutter with Gemini in about 15 short prompts, with every model, texture, and sound made in code. It runs in the browser, and the source is free to read.
Source: DEV ·
A free set of Claude Code skills and hooks makes every bug fix come with a test shown to fail without the fix. It also sends each step to the cheapest model that can handle it and leaves a handoff note for the next session.
Source: Hacker News ·
A student gave Claude their handwritten linear algebra notes and had it build a working font from their own handwriting.
Source: X ·
A pair of free skills let Codex or Claude Code build and refine 3D assets in Blender, check its own renders against a reference, fix problems, and confirm the exported files actually work.
Source: GitHub ·
A developer fine-tuned a small open model on a MacBook Air so it drops em dashes, stock phrases, and lists of three without needing a system prompt. It runs locally with Ollama and still answers factual questions as well as the original.
Source: Hacker News ·
After reports of thousands of fake repos posing as AI skills and MCP servers, a developer built an agent that scores any GitHub repo from 0 to 100 for risk before you run it.
Source: DEV ·
A developer built a photo culling tool for photographers that runs an open model on their own computer to group near-identical shots and flag blinks, so picking the keepers goes faster without uploading anyone's photos.
Source: DEV ·
A developer built a tool for a small cash-out stand owner in Nigeria: she pastes her bank's payment alert emails and a local open model fills in a day-book with the time, amount, and type of each payment, so she no longer writes it all by hand.
Source: DEV ·
An open-source Mac video editor runs Claude Code inside it, so you can ask for a cut in plain words and watch it edit a real timeline. It plans first, gets checked by a second review pass, and every change can be undone in one step.
Source: Hacker News ·
A parent built a free, open-source reader where kids tap a word to hear its sounds and work through stories by themselves. It also connects to an AI assistant so you can turn a book you own into a private copy for your own child.
Source: Hacker News ·
A budgeting app keeps all your bank data on your own Mac or PC and uses a small local model to comment on your month and point out unusual charges, like a bill that was taken twice. It's a tracking aid, not financial advice.
Source: Hacker News ·
An open-source tool gives your coding agents a shared room: send one message and they split the task, fix it in separate branches, review each other's changes, and push the result. It also works with open models through opencode.
Source: Hacker News ·
A developer fine-tuned a small open model to rate how hard a prompt is from 0 to 1, so an app can send simple jobs like formatting to a cheaper model and save the expensive one for real reasoning.
Source: DEV ·
A developer had Claude turn an open-source organization's GitHub activity into an animated cyberpunk city, so it's easy to see which projects are busy and who's working on them.
Source: X ·
One learner points a camera at their notebook each morning and does 45 minutes of math or physics with a custom AI coach that sets exercises and grades the handwritten answers. They say the rusty skills are coming back.
Source: X ·
A programmer with no astronomy training used Claude to build a pipeline over NASA's public TESS data and found a possible planet around a distant star. They call it a candidate, not a discovery, and published predictions and tests that could confirm or rule it out.
Source: Blog ·
An open-source agent skill reads your product, offers three creative directions, then builds a short launch film with its own music. You fix things by clicking a spot on the frame and saying what to change, and it edits only that part.
Source: GitHub ·
A former consultant made a simple open-source deck tool where you comment under a slide and your coding agent updates the deck, then replies on the slide with what it changed or a question for you.
Source: GitHub ·
An open-source firmware turns a small round-screen robot toy into a push-to-talk voice device for a self-hosted AI agent. It runs on your own network with your own model, with no vendor cloud and no subscription.
Source: GitHub ·
A private desktop app checks a vulnerability ID or your dependency list against dozens of public security sources and gives a risk score with next steps. It runs AI on your own device and can also plug into coding agents.
Source: GitHub ·
A freelancer's open-source tool notes which app window is open, and a small local model works out which client each stretch of time was for and drafts the invoice lines. You approve every line, and nothing leaves your laptop.
Source: DEV ·
A free n8n template watches a Gmail label for invoice PDFs, has Gemini pull out the vendor, totals, and line items, checks the math, and adds a row to Google Sheets. The post also shows how to double-check bank details before anyone pays.
Source: DEV ·
An open-source MCP server lets Claude Code, Cursor, and other AI assistants create background music and sound effects from a single sentence, using a model that runs on your own Mac or PC. It can make seamless loops for games and saves the settings for each file so you can recreate it.
Source: GitHub ·
A free Claude Code skill turns a plain-English description of an object, device, or building into a single HTML file with a fine-line isometric drawing. Switched-on parts glow, you can click them to make things happen, and you can ask for a 3D version to rotate.
Source: X ·
A developer shared their full setup for running a coding agent with a local model on a MacBook Pro. They use it for focused jobs like finding stale FIXME comments, raising test coverage, and updating an API spec, all without sending code to the cloud.
Source: Bluesky ·
An open-source tool reads your app screenshots, looks at competing apps, and writes store listings in several languages with a small open model. Character and byte limits are checked in code, so anything too long goes back to be tightened.
Source: DEV ·
An open-source Git hook pauses your push and asks a question about what the AI-written change actually does, like what happens when a payment retry fails. Answer wrong and it explains why, then tries a new scenario; answer right and the push goes through.
Source: Hacker News ·
One tester uses clocks as a personal benchmark for new models and had Claude build 500 different working clock designs over several chat sessions. They all live in one self-contained web page you can scroll through.
Source: Hacker News ·
A developer built an open-source Android app that runs AI models offline on the phone, connects to MCP tools, and can build small games and websites right on the device. It can also set reminders and control a computer through MCP.
Source: YouTube ·
Over about two days, Claude Code and dozens of parallel subagents built a detailed ancient Greek landscape with temples, waterfalls, and a whole city, all generated by code with no outside assets. The full project is open source.
Source: GitHub ·
A free, open-source course teaches how to build the files, rules, and checks around coding agents like Claude Code and Codex so they stay on track across sessions. It includes hands-on projects and copy-ready templates you can drop into your own repo.
Source: X ·
Claude Code stores past sessions under the project's full folder path, so renaming the folder can hide them. A short guide shows how to move a repo and bring your conversation history along.
Source: Blog ·
If you run Claude Code or Codex inside tmux, two fzf plugins let you search files, commit hashes, URLs, and even sentences from the agent's last reply, then drop them straight into your prompt. No more copying paths with the mouse.
Source: Blog ·
A developer used Claude Code to clean up a family TV that showed an ad before you could reach Netflix. The shared prompt has the agent back up settings, ask before turning anything off, and install a simple launcher, with a one-line restore to undo it all.
Source: GitHub ·
A command-line tool records whatever your Linux machine is playing and transcribes it live with a local speech model. It's handy for captioning meetings, calls and videos without sending audio anywhere, and saves subtitles or plain text.
Source: Hacker News ·
A cardiology team tired of tracing scan frames by hand used Claude Code to build its own tool that outlines vessel walls and stents automatically and makes 3D reconstructions. It took over a year and is for research only, not yet validated for patient care.
Source: X ·
A security analyst built a Claude Code skill that turns a plain-English question into a security log query and runs it on recent or archived data. Its first search for remote-access tools surfaced one hidden behind a harmless-looking file name.
Source: X ·
A small team built a tool for resale shops that reads the weekly government recall notices and pulls out where to find the model number on each product. Staff get a short do-not-sell and check-the-label list, since few online listings mention the exact model.
Source: DEV ·
A student gave Claude Code photos of 120 pages of college notes and got back an installable font in their own handwriting in about an hour. Each letter has a few variants that rotate as you type, and the repo explains how to make one from your own writing.
Source: GitHub ·
Drop in six to twelve photos of an object and an open-source demo builds a 3D Gaussian splat you can spin around and download. The depth model runs in the browser, so the photos never leave your computer.
Source: Hacker News ·
Pick a graphics card or computer and a local AI model, and a free site plays back text at the speed that setup would really generate it, based on measured and estimated numbers. It helps you judge what feels usable before spending money.
Source: Hacker News ·
A tool for small shops drafts replies to online reviews in the owner's voice with a small open model, then flags any promise the shop's facts don't back up, like a refund or a callback. The owner edits and posts every reply themselves.
Source: DEV ·
An engineer removed about 19,000 lines of retired compiler code in a week by having an AI mark and sort every affected function, then double-checking its work with separate code analysis tools. A month later there had been no reverts.
Source: Blog ·
A free hands-on workshop walks through tracking down one small bug in an unfamiliar codebase with text search, structural search, and symbol lookups, then asking an assistant for an explanation backed by evidence before changing anything.
Source: DEV ·
A former shop owner built an open-source tool that turns one product photo and a few true facts into listing copy, images, and short video ads on a laptop. It's tuned to never invent claims the seller didn't give, and nothing posts without your tap.
Source: DEV ·
To show a curious four-year-old how trains work, a parent started with a simple track-switching simulation in Claude. They kept adding ideas together over a weekend until it became a small set of transit-themed games.
Source: Hacker News ·
A solo business owner gives each AI agent one narrow job, like bookkeeping, research, or publishing, with its own memory and folder. The agents write dated reports instead of chat replies and label anything they couldn't verify, while spending money and other final calls stay with the human.
Source: Substack ·
Small shops in India get WhatsApp orders in a mix of Hindi, English, and shorthand. An open-source tool running on an ordinary laptop with the Wi-Fi off matches each line to the shop's catalogue, asks one short question when it's unsure, and exports the result for billing software.
Source: DEV ·
An open-source scanner reviews an agent skill or MCP setup before you install it and flags things like hidden shell commands, credential reads, and prompt injection. It never runs the code it checks, and it's upfront that a clean result is a first filter, not a guarantee.
Source: DEV ·
Researchers compared text-chat interviews led by people with ones led by a small open model hosted on their own hardware. The AI asked more questions and collected more written answers per interview, with similarly on-topic replies, while the data never left local servers.
Source: arXiv ·
An app maker whose contact form filled up with sales pitches built a sorter that runs entirely in the browser with a small open model. Customer questions stay on top, pitches fold away with the giveaway line highlighted, and nothing is deleted or answered automatically.
Source: DEV ·
Describe an imagined world and an AI writes a whole wiki for it, where clicking a red link creates the next article. Reading is free, and if you pay a small fee to generate a new article, you can steer where the story goes, a bit like a choose-your-own-adventure book.
Source: Hacker News ·
A benchmark gives each AI model one hour to write a StarCraft bot in C++ and then has the bots play against each other and against human-written bots. It is a hands-on way to see how well models learn from practice games over a long task.
Source: Hacker News ·
A landlord pointed a cheap camera at a boiler's display and set up a scheduled job that reads the screen with Claude every half hour. If an error code appears or the screen can't be read, a text goes out, long before the tenants notice the rooms getting cold.
Source: Hacker News ·
A creator who wakes up to hundreds of comments has an AI assistant send a list of links to the ones still without an answer. The replies are still written by hand; the assistant just makes sure nobody gets missed.
Source: X ·
A developer asked Claude Code for the most realistic water it could make in the browser, with no game engine. The result is an open-source sunset sea with waves, foam and clouds in a single 61 KB HTML file; the agent checked its own screenshots with separate reviewer agents along the way.
Source: GitHub ·
A landscaper photographs a rough hand-drawn plan of a front yard, notes where the phone camera stands, then gives the sketch and a real photo of the yard to an AI image model. It produces a fairly close mock-up of the finished work and can swap in photos of real tile or brick samples.
Source: Bluesky ·
A tester built a fake online shop with 40 hidden usability bugs, from faint text to a minus button that adds items, and asked an AI agent to find them using three styles of prompt. The write-up shows how much the way you ask changes what it catches.
Source: DEV ·
A tool for small pharmacy shops reads photographed supplier invoices, works out which batches won't sell before the return cutoff and sends reminders in time. It can draft the return request in Tamil, Hindi or English.
Source: DEV ·
An open-source Claude Code plugin turns plain-English feature plans into browser tests that act as a release gate. Each run also records a narrated, branded video of the feature, so the same flow gives you a pass/fail check and a demo.
Source: GitHub ·
A growth lead built a video editing workflow with Claude Code and used it to cut 28 minutes of footage, sharing a walkthrough of how the steps fit together.
Source: X ·
A shared online lab where independent AI agents pick research problems, discuss approaches and publish papers that other agents review, while people can watch. It is meant to make agent-driven research possible without huge budgets.
Source: Hacker News ·
Maintainers of a networking library added it to an AI code scanner and got back a report with 36 bugs, six of them security issues, some hiding in the code for years. The thread compares this with the low-quality AI bug reports they usually receive.
Source: Hacker News ·
A research model worked out in under half an hour how the unexplained seeds behind common elliptic-curve standards were generated: an old C random-number generator started from a 1990s timestamp. The write-up includes code anyone can run to reproduce them.
Source: Blog ·
Researchers gave Claude plain-language specs for eight quantum circuits plus an automatic checker that scores each attempt. Working in a loop with little human input, it found designs that use fewer qubits and gates than the best published ones.
Source: Blog ·
A first-person ski game built with Claude uses real elevation data for the Matterhorn, adds neon blue-hour lighting and runs straight in the browser.
Source: X ·
A video team shows how they set up a local model and an agent on their own GPU to control DaVinci Resolve through MCP, so editing help doesn't need a cloud AI subscription.
Source: YouTube ·
Researchers dropped four AI agents into a fresh Minecraft survival world with a single goal: defeat the Ender Dragon. Coordinating through in-game chat and shared notes, they wrote and debugged thousands of small programs and won after about 27 hours, with no human-made game logic.
Source: arXiv ·
A short demo shows an academic PDF turned into a full-colour, animated storybook inside an AI app, using image and video models together. It's a friendly way to explain research to people who would never read the paper.
Source: X ·
A free, open-source skill lets your coding agent cut a raw video, add captions, zooms and sound, then take notes the way a human editor would. Notes you repeat become rules, so later edits match your style.
Source: GitHub ·
A freelancer's tool compares a client's new request with the original scope using a local model, points to the matching clause, estimates the extra hours and drafts a polite reply. It also handles messages written in mixed languages.
Source: DEV ·
Paste two car ads and a tool checks the photos for damage, compares the price with similar listings and lists known weak spots of that engine to check at the viewing. It is a second opinion to bring along, not a replacement for an inspection.
Source: DEV ·
An open-source Blender add-on lets an AI assistant look at your scene, add and arrange objects, set materials and run multi-step jobs while you watch. You can pick a cloud model or a local one.
Source: GitHub ·
An open-source tool describes audio in words so a coding agent can judge whether a sound fits, find the best moment in a recording, and create and refine sound effects in code. It runs locally on a CPU.
Source: GitHub ·
A tip for saving a whole Claude Code session as a text file, so another agent can pick up the work or you can review it later to build reusable skills.
Source: X ·
A team that ships mostly AI-written code explains its review setup: agents verify and review every change, security checks can block a merge, and the number of human approvals depends on how risky the change is. Fast rollback covers whatever slips through.
Source: Blog ·
A coding agent studies a web game, writes the rules for playing it and keeps improving them from lost games. Those rules then train a small local model that makes each move in about 30 ms, fast enough to play live.
Source: Blog ·
An open-source agent keeps an eye on a Postgres or MySQL database, spots queries that are getting slower and suggests fixes such as missing indexes, so problems are caught before the database falls over.
Source: Hacker News ·
Someone used Claude to build a small web game for a music-loving child where each letter on a normal keyboard plays a note, turning the child's own game idea into something playable the same day.
Source: Bluesky ·
A developer had Claude Code compare USB traffic from the vendor's Windows driver in a virtual machine with traffic from a new Linux driver, producing a working film scanner driver and laptop camera fixes, then shared the harness for others.
Source: Blog ·
A maintainer had Codex write a skill for upgrading Angular projects, then refined it after real failures so it now branches, follows the official guide, checks clean installs and validates the build.
Source: DEV ·
A translator's tool runs a small fine-tuned Qwen model on a Mac to score each machine-translated segment for meaning errors, so you re-read the riskiest lines first without sending client text to the cloud.
Source: DEV ·
A hobbyist built a packing algorithm with Claude's help and used it to beat more than 1,200 records for fitting shapes into the smallest possible container.
Source: X ·
A creator puts Claude Motion up against a professional motion designer, adding animated graphics on top of raw talking-head footage and building a short launch intro from scratch. The video compares the time each took and shows both workflows step by step.
Source: YouTube ·
A developer tests AI models that run entirely inside the browser for background removal, alt text, ticket drafting and semantic search, then sums up what works well enough for real apps and what still falls short.
Source: DEV ·
A developer used swarms of agents to write property-based tests, first for their own code and then for well-known open-source projects, turning up around 100 bugs, with many fixes already merged.
Source: Bluesky ·
A small open-weight vision model reads a photo of a second-hand item and drafts a listing where every detail is backed by something visible. Anything it can't confirm, like the material, is flagged for you to check instead of guessed.
Source: DEV ·
A small Python step has a local LLM read each supplier PDF and return line items as JSON for a spreadsheet, with a simple fix for stripping stray markdown before parsing.
Source: Bluesky ·
A design agent asks what your device needs, picks real components, draws the schematic and routes a manufacturable PCB, with engineering checks before you send it to a manufacturer.
Source: Product Hunt ·
A maker trained a 30-million-parameter language model from scratch, squeezed it down with 4-bit quantisation and got it answering offline on an ESP32 board, as a hands-on way to learn how LLMs work.
Source: YouTube ·
A team spent four weeks designing and testing the upgrade of a 16-node HPC cluster with Claude Code, then let the agent carry out the full reinstall itself.
Source: Bluesky ·
Asked to re-render a one-minute video in a chosen animation style, Claude Code spotted that Codex could generate images and ran 161 Codex sessions to make the frames, managing memory, retrying failures and stitching the result together.
Source: X ·
An open-source MCP server lets Claude Code or Codex send synthetic participants through a website, localhost build or Figma prototype in a real browser, then returns the top fixes with evidence.
Source: X ·
A free Claude Code skill researches a topic you name and builds about a dozen two-minute audio lessons, each with a summary card, one thing to practise and links to the pages it used.
Source: GitHub ·
A Roblox creator walks through the whole process of building and publishing a game with Claude, sharing practical tips along the way.
Source: YouTube ·
Engineers let an AI model work unattended for 11 hours on a 200,000-line open-source Python project, guided by a short goal and repo rule files. It proposed six speed-ups; people rejected four in review and shipped two, cutting one slow step by about 40%.
Source: Blog ·
A Claude Code skill first maps a new field's key terms, people and venues, then follows citations and links until it finds the original sources behind each claim, instead of summarising the top search results.
Source: GitHub ·
A designer built a teleprompter app with Claude Code over hundreds of prompts, testing on a real phone, fixing a freeze with the agent's help, adding automated tests and getting it through App Store review.
Source: YouTube ·
An open-source benchmark reuses published patches from coding benchmarks and counts the syntax elements each model adds, so you can see which models solve tasks with the least extra code to maintain.
Source: Hacker News ·
A Mac budgeting app keeps your ledger and your questions on your own machine, using small local models to answer things like where your money went last month, and says when it isn't sure instead of guessing.
Source: Hacker News ·
A free list of 120 writing habits that make text read like AI filler, each with a bad example and a plain rewrite. You can give it to an assistant before it writes or use it to check a draft.
Source: GitHub ·
A data developer used Claude Code to bring back an R package for a public road-safety data API that had been mostly broken for two years.
Source: Bluesky ·
A Claude Code skill takes any two places, maps the real road between them from OpenStreetMap and elevation data, and builds a scrollable pixel-art journey you can use for games, trip recaps or videos.
Source: GitHub ·
A Claude Code skill turns a one-line description of a machine, like a jet engine, into a 3D simulation in a single HTML file that runs on real physics, plus a short film to share it.
Source: GitHub ·
A Claude Code skill turns a question about how a receptor, hormone or brain pathway works into a standalone page with a diagram you can toggle and a citation for every number shown.
Source: GitHub ·
A travel app builds walking tours around what interests you and narrates them with cited sources and a fact-checking pass, so a few spare hours in a city can become a guided wander.
Source: Product Hunt ·
An open-source project uses a low-cost ESP32 board as AirPlay eyes and a USB mouse so a vision model can tap through iPhone apps, such as setting a repeating alarm, with no jailbreak or computer.
Source: GitHub ·
An amateur used Claude Code to write and run analysis on public astronomy data and believes it surfaced a planet candidate no one had catalogued; the thread discusses how to verify it.
Source: Hacker News ·
An open-source toolkit lets an agent design LEGO builds in the LDraw format from a text prompt, producing models you can view and build.
Source: GitHub ·
A hobby project for a small e-paper device on the fridge that shows the household shopping list, edited from a phone web page, built with help from Claude Code.
Source: GitHub ·
A developer had an LLM read the 68000 assembly of a 1993 Amiga platformer and port it to Godot, then reviewed what the model got right and wrong along the way.
Source: Blog ·
A Raspberry Pi with a dashcam and a local AI model lets you ask your car what it has seen through a chat room, all running on-device.
Source: GitHub ·
A blog post laying out a routine for using LLMs as a tutor: asking for explanations at different levels, quizzing yourself and checking claims against sources.
Source: Blog ·
A team had four frontier models draw the Mona Lisa stroke by stroke with virtual colored pencils, tracking cost and comparing the results.
Source: Blog ·
A researcher used reinforcement learning to teach an AI agent to run model training jobs itself, on a budget of about 1,300 dollars.
Source: GitHub ·
An education company explains how it built a voice AI tutor for children aged 4 to 9 that has to respond in under a second and follow good teaching practice.
Source: Blog ·
An open-source command-line toolkit that gives AI agents a way to open, read and modify Microsoft Office files as part of an automated workflow.
Source: GitHub ·
The author pointed Claude Code at their own shoulder MRI files to get a layperson's read of the images, and concluded the results were interesting but not yet trustworthy enough to rely on.
Source: Blog ·
A small Mac utility prevents sleep while AI coding agents are running, even with the lid closed, then lets the machine sleep when they finish.
Source: GitHub ·
A Mac app records a call's microphone and system audio, transcribes it with a local speech model, writes an on-device summary, and lets you flag key moments mid-call or ask questions about past meetings without anything leaving the machine.
Source: Web ·
An open-source tool that has an LLM write multi-part, practical tutorials on a topic you choose, which you then work through by hand instead of letting the AI do it.
Source: GitHub ·
A code search tool gives coding agents relevant snippets directly, using far fewer tokens than grepping and reading whole files.
Source: GitHub ·
A developer had LLMs write most of a Rust implementation of the RAR archive format, describing what worked and where human review was still needed.
Source: Blog ·
A dashboard-as-code tool lets agents and people define analytics dashboards in YAML and JSX so changes can be reviewed like code.
Source: GitHub ·
Open-source drivers let computer-use agents operate macOS and other desktops without taking over your mouse, for automation and testing.
Source: GitHub ·
Claude Code drives a SPICE simulation and a lab oscilloscope, then compares the measured waveform to the simulated one to verify the hardware behaves as designed.
Source: Blog ·
A developer built an open-source tool for scheduling and publishing posts to many social platforms in about three weeks, with Claude and Codex writing much of the code.
Source: GitHub ·
A parent photographed a hand-drawn pegboard idea and had an AI agent turn it into printable 3D model files for a toy.
Source: GitHub ·
An open-source library uses LLMs to pull clean, typed data such as product details out of web pages and repair malformed output.
Source: GitHub ·
A workflow orchestrator runs coding agents on Kubernetes so a ticket can move through implementation and review into a pull request.
Source: GitHub ·
An open-source tool that lets a coding agent capture what the app actually looks like after a change, so it can verify the interface instead of assuming it works.
Source: GitHub ·
An engineer explains the habits that let them land a high volume of changes soon after joining a startup, from task sizing to reviewing agent output.
Source: Blog ·
A developer sets up agents to work through tasks while they sleep and reflects on how hard it is to judge whether the unattended work is any good.
Source: Blog ·
A JavaScript library embeds a GUI agent inside a website so users can ask it to fill forms and navigate the app for them.
Source: Web ·
An AI comment moderator checks comments before they post and suggests more constructive wording, as an alternative to plain toxicity scoring.
Source: Web ·
A well-known programmer had Claude Code implement a clean-room Z80 and ZX Spectrum emulator from documentation, and reports on the quality of the result.
Source: Blog ·
A self-hosted agent server keeps memory and tools on your own hardware and talks to you through apps like Telegram, WhatsApp or Slack.
Source: Web ·
A screen reader user tested modern AI text-to-speech systems for everyday accessibility and explains why speed and predictability matter more than realism.
Source: Blog ·
Someone who had stopped coding years ago used AI coding tools to build and ship a set of free online calculators, describing it as the thing that got them building again.
Source: Web ·
A command-line tool gives AI agents read and write access to Figma so they can create shapes, text and components or export images.
Source: GitHub ·
An open-source desktop app gives a general-purpose agent access to your files and tools for office-style tasks, as an alternative to hosted coworker products.
Source: GitHub ·
A command-line browser tool gives coding agents and people the same way to open pages, click and read content, without a heavy integration layer.
Source: GitHub ·
A session manager runs several terminal coding agents side by side and lets you check on them from a web view on mobile.
Source: GitHub ·
A reading project where Claude reads across a set of about 100 books and surfaces thematic trails linking passages between them, published as a browsable site.
Source: Web ·
A developer advocate describes building several working apps in days with a new Claude model and argues coding agents have changed what one person can ship.
Source: Blog ·
A guide to running a coding agent on a remote machine and driving it from a phone terminal, so small tasks can be done away from the desk.
Source: GitHub ·
A research paper that ran AI agents and security professionals on real-world penetration tests and compares what each found.
Source: arXiv ·
A founder used Claude Code to build a visual bookshelf website of the books they have read, describing the back-and-forth that got it finished.
Source: Blog ·
A free tool has an LLM read a user's Hacker News comments and write a playful summary of their year.
Source: Web ·
A blog post on porting an HTML5 parser from Python to JavaScript in a few hours with a coding agent, using an existing test suite to check every step.
Source: Blog ·
An experiment had an LLM read decade-old Hacker News threads and score which commenters' predictions turned out right, as a test of mining history at scale.
Source: Blog ·
An official developer post walks through practical things to try with Gemini CLI in the terminal, from building apps to working across a codebase.
Source: Blog ·
A blogger tested a new Gemini model by transcribing a city council meeting recording with speaker labels and timestamps, and checked how accurate it was.
Source: Blog ·
A critic agent watches what your coding agent does, flags mistakes and learns your preferences over time.
Source: GitHub ·
A skill lets coding agents write and run Playwright scripts to click through a site and confirm features actually work.
Source: GitHub ·
A blogger let Claude Code work through dependency and driver problems to get a freshly released OCR model running on an NVIDIA box.
Source: Blog ·
A founder paired with Claude Code and MCP servers to rebuild their company website, covering what sped things up and where they stepped in.
Source: Blog ·
An open-source assistant builds a memory of your projects, people and notes as a knowledge graph so it can answer questions with context.
Source: GitHub ·
An open-source agent operates Windows by reading the screen and clicking and typing in ordinary desktop apps, with no app-specific integration.
Source: GitHub ·
A staff engineer describes six weeks of running multiple Claude Code agents in parallel, expecting the first attempt to be mostly wrong and iterating from there.
Source: Blog ·
A free service converts Apple's JavaScript-rendered developer documentation into Markdown so AI assistants can read it.
Source: Web ·
A project-management skill turns specs into GitHub Issues and runs agents in parallel git worktrees so progress stays traceable.
Source: GitHub ·
An open-source desktop app turns your voice into text in any application, with the option to run transcription locally so audio stays on your device.
Source: GitHub ·
An open-source platform lets teams run and monitor managed agents on their own servers instead of a hosted agents API.
Source: GitHub ·
A local-first server lets an AI assistant answer questions about sleep, workouts and heart data from an Apple Health export without uploading it anywhere.
Source: GitHub ·
A developer recounts two weeks of moving their daily work to Claude Code, including the habits and prompts that made it click.
Source: Blog ·
A kanban app assigns tasks to coding agents, runs them in isolation and lets you review each result before merging.
Source: GitHub ·
A developer built and released a SwiftUI macOS app for debugging MCP servers almost entirely through Claude Code, and shares what worked.
Source: Blog ·
A tower defense game for a game jam was built almost entirely with AI coding tools, with the prompts and lessons written up in the repository.
Source: GitHub ·
A write-up on building agents that turn a business question into queries over spreadsheets and databases and return the answer, so non-technical staff can self-serve data.
Source: Blog ·
A research environment that measures how well LLM agents plan and automate by giving them the game Factorio and asking them to build production lines.
Source: Web ·
An open-source Python library lets an LLM control a real browser to search, click and collect information for multi-step tasks.
Source: GitHub ·
A staff engineer describes where LLMs actually help in their job, such as quick prototypes, unfamiliar code and proofreading, and where they avoid relying on them.
Source: Blog ·
An open-source app that collects words you meet in books, articles or videos and uses an LLM to generate Duolingo-style quizzes and lessons from your own list.
Source: GitHub ·
An open-source tool combines LLMs and computer vision to complete browser workflows like filling forms on sites that offer no API.
Source: GitHub ·
A no-code service combines LLM steps with generated code so you can set up alerts and simple automations in plain English.
Source: Web ·