Grok Build Guide

Grok Build is SpaceXAI’s terminal coding agent. It entered early beta in May 2026, and it has been open source under the Apache 2.0 license since July 14, 2026.

Unlike a coding chatbot, it can:

• Understand your codebase
• Edit files
• Run shell commands
• Search the web
• Plan before building
• Hand work to parallel subagents
• Run headless in scripts and CI
• Work inside editors via the Agent Client Protocol

As a coding agent, it changes your files. Treat it like a new teammate with access to your repository.

Grok Build opened to SuperGrok and X Premium+ subscribers on May 25, 2026. It also works with an xAI API key. See x.ai/cli for current plans.

The docs name Grok 4.7 as the latest model. Change models with /model, or add custom ones in the config file.

Confirm the install with grok --version. The first launch opens your browser so you can sign in. You can also set an API key:

export XAI_API_KEY="your-key"

Keep API keys out of your repository, always.

Plan

Grok Build drafts a plan and edits only the plan file. Enter it with /plan plus a description.

Auto

A classifier auto-approves safe tools. Enter it with /auto.

Always-approve

No permission prompts at all. Enter it with /always-approve.

Cycle through the modes with Shift+Tab.

Use Plan for anything that matters. Save Always-approve for a disposable sandbox.

In Plan mode you can approve the plan, comment on single steps, or rewrite it before anything runs. Make sure the plan:

• Solves the real problem
• Touches only the right files
• Includes tests
• Adds no unneeded dependencies
• Explains anything risky

Review it with /view-plan, then let the build start.

Grok Build reads AGENTS.md files by default. Put one at your project root covering:

• What the project does
• The tech stack
• How to install, run, test, and build
• Code style rules
• Folders it must leave alone
• Commands that need your approval first

• Request one feature or fix at a time
• Run the tests after every change
• Try the result yourself
• Commit to Git once it works
• Open a fresh session with /new when you switch tasks

Small steps that are tested and committed keep every mistake easy to undo.

Grok Build can split larger tasks across parallel subagents, and Git worktree support keeps each one’s changes separate.

/workflow

Runs a saved workflow

/create-workflow

Creates a new workflow

/deep-research

Runs a research workflow on one question

Read every subagent’s changes before you merge.

• Skills: invokable ones show up as slash commands
• Plugins and marketplaces: packaged features
• Hooks: your own checks at set moments
• MCP servers: connections to other tools and data

Install extensions only from trusted sources. Each one sees everything the agent sees.

Add the -p flag to run Grok Build in scripts and automations, with streaming JSON output when you need it.

• Generate release notes
• Triage new issues
• Fix lint errors
• Write tests for changed files

In automations, grant the minimum access it needs and review each change before it reaches production.

With the Agent Client Protocol, Grok Build can run inside ACP-compatible editors and agent apps, so you are not tied to the terminal.

On July 12, 2026, a security researcher showed that Grok Build version 0.2.93 uploaded whole Git repositories, history and secrets included, to cloud storage controlled by xAI. SpaceXAI switched off the upload server-side around July 13 and open-sourced Grok Build a day later.

With any coding agent, protect yourself:

• Keep secrets in environment variables, never in code or Git history
• Rotate every key that was ever committed
• Give agent work its own test key
• Check your data and training settings
• Stick to your company’s approved tools for client code

/new or /clear

Start a fresh session

/sessions

Manage sessions

/model <name>

Switch models

/memory

Browse memories

/settings

Open settings

/imagine <prompt>

Generate images

/feedback

Send issues and ideas to the team

/quit

Exit

Act as a senior engineer on this project. Read AGENTS.md before anything else. Task: [feature or bug]. Begin in plan mode: list the files you will change, your approach, the tests you will add, and any risks. Wait for my approval. Then build one step at a time. Run the tests after every step and show me the results. Ask before you add dependencies, touch [protected folders], or run destructive commands. Never print or commit secrets.

• Running Always-approve on your main machine
• Pointing it at repositories that hold secrets
• Merging code you have not read or tested
• Installing plugins or MCP servers you do not trust
• Skipping Git commits between steps
• Pasting API keys into the chat