Grok Build

Short verdict: Grok Build is xAI's terminal coding agent: a full-screen TUI that plans, edits files, runs commands and splits big jobs across parallel subagents, on Grok 4.7 by default. Plan mode, workflows and support for Claude Code instruction files are well documented, and the CLI has been Apache-2.0 open source since July 2026. The catch is the bundle: it comes with The X Subscription, whose weekly allowance, of unpublished size, is shared with Grok chat, image and video generation, Grok Bot and Cursor. It suits people who already pay for Grok or Cursor or want an open agent for their own models, not anyone who needs published limits.

How we reviewed it: We did not sign up for Grok or run Grok Build against a model. On 21 October 2026 we read the Grok Build page, docs and changelog, the pricing page and The X Subscription docs, xAI's announcements since May 2026, its consumer FAQ and privacy policy, and the GitHub repository, and installed version 1.0.62 from npm only to read grok --help and its privacy notice. Plans and limits change often; check the pricing page before you pay.

What Grok Build costs and includes

What sets Grok Build apart

Plans you approve step by step

In plan mode only the plan file can change: xAI says every edit is blocked until you approve, and you can approve the plan, comment on single steps or rewrite it. Outside plan mode, permissions range from asking for each tool call, through auto mode (a classifier approves safe tools), to always-approve.

Subagents, workflows and goals for parallel work

Subagents come in three built-in types (general-purpose, explore and plan), each with its own context window, and can run in separate git worktrees. Workflows, added on 23 July 2026, fan a task out across what xAI calls hundreds of parallel agents, verify the results and report once. /goal keeps working until a panel of reviewer subagents agrees the objective is met.

It reads your Claude Code setup

Grok Build reads AGENTS.md, CLAUDE.md and CLAUDE.local.md, plus .claude/rules/ and .cursor/rules/, and /import-claude brings settings over. Company deployments can keep Claude Code's managed-settings.json permission rules and MCP allowlists, while Grok's own /etc/grok/requirements.toml pins policies such as sandbox profiles. Skills, hooks, MCP servers and a plugin marketplace work too.

Open source, with any model you configure

xAI published the agent loop, tools, terminal UI and extension system under Apache-2.0 on 15 July 2026, and says Grok Build can run fully local-first against your own inference. Custom models go in ~/.grok/config.toml, an xAI API key works for scripts and CI, and headless mode (grok -p) returns plain text or JSON, as the docs explain.

Limits and costs to watch

• No published allowance. Free is "Limited" and paid plans are multiples of Plus's "Base" pool, which is shared with chat, media, Grok Bot and Cursor: a week of video work leaves less for coding.

• Overage costs more. When the pool runs out, paid features pause until the weekly reset unless you buy Extra Usage Credits (from $5), which xAI says cost more per action than included usage.

• Bundle only. Grok Build comes inside The X Subscription or with a pay-as-you-go API key. Existing SuperGrok, Cursor and X Premium plans continue, but moving one to The X Subscription is permanent.

• Loose safety defaults. The sandbox (Landlock on Linux, Seatbelt on macOS) is off by default, and always-approve is one flag away.

• Young and fast-moving. Early beta on 25 May 2026, version 1.0.0 on 7 August, new versions most days in the changelog; the docs lag (/share, removed on 9 October, is still documented).

• Two privacy defaults. The CLI's training setting stays off unless you opt in, but xAI may train on Grok chat conversations unless you turn off "Improve the model".

We could not verify the size of any allowance, whether legacy SuperGrok and X Premium+ plans keep their own limits, or press reports that SuperGrok Heavy subscribers got Grok Build before xAI's 25 May announcement.

Pros and cons

Pros

• Usable on the free plan, and included from $8 a month with Grok chat and Cursor.

• Plan mode blocks edits until you approve; subagents, workflows and /goal handle parallel and long-running work.

• Apache-2.0 source, custom and local models, and xAI API keys.

• Reads Claude Code instruction files and policies, so moving over costs little.

Cons

• No published allowance, and the weekly pool is shared with chat, media and Cursor.

• Sold only in a bundle; moving a legacy plan into it cannot be undone.

• The sandbox is off by default.

• Docs trail releases; the GitHub mirror lags and takes no contributions.

• No IDE extension or desktop app; editors need ACP.

How Grok Build compares

• Claude Code has no free tier (Pro is $20 a month, or $17 billed annually), uses 5-hour plus weekly limits shared with Claude chat, and is closed source. Choose it for Claude models and IDE, desktop and web versions of the same agent.

• OpenAI Codex also has an open-source CLI (Apache-2.0, 130.3k stars) that runs local models; full Codex comes with ChatGPT Plus at $20 a month. Choose it if you already pay for ChatGPT or want cloud tasks that keep running while your computer sleeps.

• Cursor shares the X Subscription and its weekly pool: Plus at $8 a month gives its editor agent with Grok models only, and Premium at $30 adds Cloud Agents, Bugbot reviews and third-party models at API prices. Choose Cursor for an editor and cloud agents; one subscription covers both.

For other agents and editors, see all AI coding tools; for prompt-to-app tools, see AI app builders.

Who should use Grok Build

Grok Build fits developers who already pay for Grok, X Premium or Cursor and want a terminal agent at no extra cost, Claude Code users who want their instruction files to carry over, and anyone who wants an Apache-2.0 agent to point at other models. Look elsewhere if you need published usage numbers, an IDE extension or a long track record: it is five months old, and its allowance shrinks whenever you use Grok for anything else.

FAQ

Is Grok Build free?

Yes, in a limited form: as of 21 October 2026 the Free plan of The X Subscription includes Grok Build marked "Limited", and xAI does not publish how much that covers. Paid plans start with Plus at $8 a month ($80 a year). You can also run it with an xAI API key and pay per token, for example $2 per million input tokens and $6 per million output tokens for Grok 4.7.

How much does Grok Build cost?

Grok Build is not sold on its own; as of 21 October 2026 it comes with The X Subscription: Plus $8 a month, Premium $30, Super $100 and Ultra $200, or $80, $300, $1,000 and $2,000 a year. Teams pay $30, $100 or $200 per seat. Every plan's weekly usage pool is shared across Grok, Grok Build and Cursor, and Premium, Super and Ultra get 4, 15 and 50 times the Plus allowance.

Does Grok Build train on my code?

Not by default: the CLI's setting for sharing coding data for training is off unless you opt in, according to the notice in version 1.0.62 and its open-source code. Grok chat works differently, since xAI may train on consumer conversations unless you turn off "Improve the model". Team admins can also enable zero data retention for Grok Build.

Is Grok Build open source?

Yes. xAI released the CLI, agent loop and terminal UI under the Apache-2.0 license on 15 July 2026 at github.com/xai-org/grok-build, which had 27.5k stars on 21 October 2026. The repository is a periodic mirror of xAI's internal code that does not accept outside contributions, and the Grok models it uses by default run on xAI's servers.