Grok Build: xAI's Coding Agent, Free Access, Models, and Privacy

Grok Build is xAI's coding agent. You install it as the grok command and run it in a terminal against a local repository, where it reads code, edits files, runs commands, and asks before doing anything risky. It also runs headless (grok -p) for scripts and CI, and inside other apps through the Agent Client Protocol (ACP). xAI now publishes as SpaceXAI.

Is it free? As of September 26, 2026, x.ai/build says Grok Build is "available to try for Free," and the x.ai pricing table lists Grok Build on every plan from Free to Enterprise. That's a change from the May launch, when it was limited to SuperGrok and X Premium+ subscribers. One of xAI's own pages, grok.com/build, still says that (details below).

Which model does it use? xAI says Grok Build is now powered by Grok 4.7, released September 21, 2026. The API model named grok-build-0.1 is a separate, cheaper coding model. You can also point Grok Build at your own or a local model.

Where does it run? macOS, Linux, Windows, and WSL, per xAI's install docs.

Is it open source? The client is. xAI published the Rust source on GitHub under Apache 2.0 in July. The default models and service still come from xAI, and the repository doesn't accept outside pull requests.

Does it read Claude Code config? Yes. xAI documents that it reads CLAUDE.md, .claude/rules/, and Claude Code skills, plugins, hooks, and MCP servers with no setup.

Privacy: In July, a researcher reported that Grok Build uploaded entire git repositories to xAI storage. xAI switched that off server-side. Before you point it at private code, check the /privacy setting and, on a team plan, whether zero data retention is on.

It isn't Grok Bot. Grok Bot is xAI's AI-teammate app that works across business tools on a shared cloud computer. Grok Build works in your codebase.

Which Grok Build?

xAI uses the name for three different things, and search results mix them up.

Grok Build, the coding agent. The grok CLI/TUI this page covers. It runs on your machine, in a terminal, headless, or over ACP. As of September 26, 2026, x.ai/build says it's free to try.
Grok Build, the app builder. The in-chat builder that grew out of Build Mode. You describe an app, game, site, or dashboard on grok.com or in the Grok iOS and Android apps, and Grok builds it and publishes it to a grok.me link. xAI's August 19 post says it's on every plan.
grok-build-0.1. An API model ID for agentic coding, sold per token ($1 input and $2 output per million tokens) on the xAI API, OpenRouter, and other gateways.

The app builder launched on July 28 as Build Mode, limited to SuperGrok Heavy. xAI's August 19 post, "Grok Build on web and mobile," renamed it and opened it to every plan. That post is about the app builder, not the terminal agent. If you've read that "Grok Build is on every plan," check which one the source means. The two are connected in one place: the app builder can export a project to GitHub so you can keep working "in Grok Build for the terminal."

grok-build-0.1 is a model, not a product. How it relates to the agent is covered under "Which model Grok Build uses" below.

How the coding agent works

Official docs

What xAI documents

Three ways to run it. An interactive, mouse-aware full-screen TUI; headless with grok -p "prompt", with plain, json, or streaming-json output; or as an ACP agent over stdio with grok agent stdio, for editors and your own tools.
Sign-in. On first launch, Grok opens a browser to sign in to your Grok account. Device-code login covers SSH and containers. An XAI_API_KEY works for scripts and CI.
Permissions. By default it asks before running tools. There's an auto mode that uses a classifier and an always-approve mode, plus allow and deny rules. Sandbox profiles are optional and off by default. The settings reference says the first permission prompt in a session preselects the always_allow_all_sessions option by default, so read it before you press Enter.
Plan mode. Only the plan file can be edited until you approve the plan.
Subagents and worktrees. Parallel child sessions with their own context. Built-in types are general-purpose, read-only explore, and plan. Worktree sessions run in an isolated git checkout under ~/.grok/worktrees/.
Skills, plugins, hooks, MCP. Skills are folders with a SKILL.md that also appear as slash commands. Plugins bundle skills, agents, hooks, MCP servers, and LSP servers, and there's a built-in marketplace. Hooks run on lifecycle events, and PreToolUse is the only one that can block.
Workflows. Saved orchestration scripts that fan work out to subagents and verify the results, plus a built-in /deep-research workflow.

Two hook details matter if you rely on hooks for safety. Hooks are fail-open: per xAI's docs, a hook that times out, crashes, or returns malformed output doesn't block the tool call. Only an explicit deny does. And project hooks don't run until you trust the folder with /hooks-trust or --trust.

Access, plans, and the free tier

xAI isn't consistent about this. grok.com/build, the page that ranks second for "grok build," still says Grok Build is "early access for SuperGrok and X Premium+ subscribers." That matches the May 25 launch post, not the current product page, pricing page, or API docs. Guides that say you need SuperGrok or X Premium+ to use Grok Build are repeating the launch terms.

The current pricing page doesn't list X Premium+ as a plan. We can't confirm from xAI's pages whether X Premium+ still includes Grok Build.

Usage limits

xAI doesn't publish a numeric Grok Build quota for any plan. Not in messages, tokens, tasks, or hours. Pages that give one are estimating.

Inside the CLI, /usage shows credit usage and billing. According to Grok Build's changelog, it's hidden for free and X Basic personal accounts (July 4), for API-key sign-in, and for enterprise setups.

xAI hasn't documented how free-tier Grok Build limits are measured or when they reset. Some users report hitting a "free Grok Build usage limit" message with no reset time shown. We haven't seen it ourselves.

Which model Grok Build uses

This is where the most confusion is. xAI's sources point in more than one direction, so here's what each says.

Official docs

What xAI's sources say, as of September 26, 2026

Grok 4.7 is the default. x.ai/build says Grok Build is "now powered by Grok 4.7." xAI's Grok 4.7 docs list Grok Build as "the default model of the coding agent." Changelog 1.0.40 (September 20) reads "Grok 4.7 has arrived!"
The settings docs recommend default = "grok-build". The example config.toml sets [models] default = "grok-build" with the comment "recommended for coding / agent sessions." The docs don't say which underlying model grok-build resolves to.
grok-build is also an agent name. The settings reference lists GROK_AGENT, whose default is grok-build, as the built-in agent profile. That's the harness configuration, not a model.
grok-build-0.1 is an API model. xAI released it in public beta on May 29 and called it "the same model that powers Grok Build" at the time. Its docs list a 256,000-token context window, $1 input and $2 output per million tokens, and grok-code-fast-1 as an alias.
Some pages are behind. The x.ai/pricing table still lists Grok 4.6 as the plan model, and the terminal mockups on x.ai/build show grok-4.6. Grok Build's changelog set Grok 4.5 as the default on July 18.

Custom and local models. Grok Build accepts any OpenAI-compatible or Anthropic Messages-compatible endpoint. You add a [model.<id>] block to ~/.grok/config.toml and set it as the default, or pick it with -m or /model. xAI's open-source post says you can compile the client and point it at local inference. Subagents can be routed to different models.

Install and start

xAI's documented install commands:

curl -fsSL https://x.ai/cli/install.sh | bash

That covers macOS, Linux, WSL, and Git Bash. On Windows PowerShell, use irm https://x.ai/cli/install.ps1 | iex. xAI's enterprise docs also list npm install -g @xai-official/grok.

Then run grok inside a project folder and sign in. xAI's Grok Build docs cover setup, config, and authentication. We haven't run the installer.

Claude Code compatibility

This is the most useful detail if you already run Claude Code, or run several coding agents on one repository.

Official docs

What xAI documents

Instructions: reads AGENTS.md, CLAUDE.md, Claude.md, CLAUDE.local.md, and .claude/rules/*.md, plus .cursor/rules/. Gitignored files are skipped.
Extensions: reads Claude Code marketplaces, plugins, skills, MCP servers, agents, and hooks from .claude/settings.json. Claude tool names such as Bash, Read, and Edit are mapped to Grok's automatically. Cursor's equivalents are read too.
Sessions: the session picker can resume recent Claude Code, Codex, and Cursor sessions (changelog, July 13). /import-claude opens a settings import.
Managed settings: Grok reads a subset of Claude Code's managed-settings.json: permission rules, MCP allowlists, a few telemetry flags, and marketplace restrictions.
Switches: each compatibility scanner is on by default and can be turned off under [compat.claude] and [compat.cursor] or with environment variables.

One exception matters for security teams. xAI's enterprise docs say Claude Code's disableBypassPermissionsMode is not applied to Grok's always-approve mode. If your organization locks that down for Claude Code, a developer can still turn on --yolo in Grok unless you set disable_bypass_permissions_mode = true in Grok's own root-owned /etc/grok/requirements.toml.

Open-source status

Privacy and repository-upload controls

disable_codebase_upload isn't a setting you can change. It was a server-side flag. We didn't find it in xAI's Grok Build docs or config reference. The controls you do have:

Official docs

Current controls xAI documents

/privacy opens the "Coding data, retention, and training" setting. In the public source, it sets a coding-data-retention opt-out on your account.
Zero data retention is enforced per team. For ZDR teams and enterprises, xAI says no prompts, code, or responses are kept at the inference layer.
Network: code.grok.com handles remote session sync and sharing. xAI's enterprise docs say that if you block it, sessions stay local-only. Only cli-chat-proxy.grok.com and auth.x.ai are required.
Policy pinning: organizations can pin settings, including turning off telemetry, in /etc/grok/requirements.toml, which user config can't override.
What the agent reads: [tools] respect_gitignore defaults to false, so search and read tools can open gitignored files unless you turn it on. Sandbox profiles can deny paths such as **/.env at the kernel level. web_fetch is off by default.

Where it fits in a coding-agent workflow

Grok Build covers the same ground as other terminal coding agents: repository inspection, planning, bounded edits, and running your tests. It doesn't change the need for a factual check on what the agent writes.

The production coding-agent workflow: scope, plan approval, bounded edits, deterministic checks, and a separate factual QA pass

workflow

We haven't compared Grok Build with Claude Code, Codex, or Cursor, and xAI's published benchmarks are for Grok 4.7, the model, not for the agent. We'll publish a comparison when we have our own test results.

Muse Code: another terminal coding agent, Meta's, with its plans, models, and platform support

entity

Grok Build vs Grok Bot

Grok Build is a coding agent. It runs on your machine, in your repository, signed in with your Grok account or an xAI API key.

Grok Bot is xAI's app for AI teammates that do multi-step work across business tools and websites. It runs on a shared cloud computer hosted by Cursor and is billed through Cursor. Most of its example jobs are operations work like sales outreach, inbox triage, and invoices.

If the finished output is a pull request, that's Grok Build's job. If it's an updated CRM or a cleared inbox, that's Grok Bot's.

Grok Bot: xAI's AI teammates, the shared cloud computer, platforms, and who can access it

entity

What we haven't verified

Whether the coding agent works on a Free account in practice, and how far free usage goes.
What the grok-build model ID in xAI's config example resolves to.
Whether X Premium+ still includes Grok Build.
Whether the repository-upload path is still switched off for every account, and what data from before July 13 was deleted.
How Grok Build performs against other coding agents. We have no first-party test.

Change log

May 25, 2026: xAI launches Grok Build as an early beta for SuperGrok and X Premium+ subscribers.
May 29, 2026: grok-build-0.1 goes into public beta on the xAI API, described as the model powering Grok Build.
July 13–14, 2026: A researcher reports whole-repository uploads. xAI switches off the upload server-side.
July 15, 2026: xAI publishes the Grok Build source under Apache 2.0.
July 18, 2026: The changelog sets Grok 4.5 as the default model.
July 28, 2026: Build Mode, the in-chat app builder, launches for SuperGrok Heavy.
August 19, 2026: The app builder is renamed Grok Build and opened to every plan.
September 20–21, 2026: Grok 4.7 arrives in Grok Build (changelog 1.0.40), and xAI announces it as the default.
As of September 26, 2026: x.ai/build and x.ai/pricing say the coding agent is free to try. grok.com/build still says SuperGrok and X Premium+ only, and the pricing table still lists Grok 4.6.

How we label public facts, community reports, and inference on this site

concept

Sources & Documentation

xAI: Grok Build product page — "Available to try for Free," Grok 4.7 as the model, install command, and feature list. Its terminal mockups still show grok-4.6.
xAI: Pricing — Plan comparison listing Grok Build on Free through Enterprise, and Grok 4.6 as the plan model.
Grok: Grok Build landing page — Still says early access for SuperGrok and X Premium+ subscribers.
xAI: Introducing Grok Build (May 25, 2026) — Launch terms: early beta for SuperGrok and X Premium+.
xAI: Grok Build 0.1 on API (May 29, 2026) — grok-build-0.1 release, pricing, and "the same model that powers Grok Build" at the time.
xAI: Grok Build is Now Open Source (July 15, 2026) — What the published source includes, and local-first use.
xAI: Introducing Build Mode (July 28, 2026) — The in-chat app builder's launch for SuperGrok Heavy.
xAI: Grok Build on web and mobile (August 19, 2026) — The app builder renamed Grok Build and opened to every plan, with grok.me publishing.
xAI: Introducing Grok 4.7 (September 21, 2026) — Availability in Grok Build and "Try it in Grok Build for free."
xAI Docs: Grok Build overview — Install, sign-in, headless use, ACP, and custom models.
xAI Docs: Grok Build settings and reference — The default = "grok-build" example, GROK_AGENT, respect_gitignore, sandbox, and compatibility switches.
xAI Docs: Skills, plugins, and marketplaces — Claude Code and AGENTS.md compatibility.
xAI Docs: Hooks — Events, fail-open behavior, and Claude and Cursor hook files.
xAI Docs: Modes and commands — Plan, auto, and always-approve modes, /usage, /privacy, and workflows.
xAI Docs: Enterprise deployments — Network hosts, authentication, managed settings, Claude Code managed-settings compatibility, and ZDR.
xAI Docs: Grok 4.7 — Grok 4.7 as the coding agent's default model, and Grok 4.7 Fast excluded from the free tier.
xAI Docs: Grok Build 0.1 — Context window, pricing, and aliases.
xAI Docs: Pricing — API token pricing and Grok 4.7 Fast pricing for Cursor and Grok Build.
xAI Docs: Grok FAQ — Shared weekly usage pool, Settings → Usage, and Extra Usage Credits.
xAI: Grok Build changelog — Release notes, including 1.0.40 (Grok 4.7) and the July session-resume and rate-limit changes.
GitHub: xai-org/grok-build — Apache 2.0 source, README, CONTRIBUTING (no external PRs), CHANGELOG, and the /privacy command source.
The Hacker News: Grok Build uploaded entire git repositories (July 14, 2026) — Third-party report of the repository upload and xAI's server-side shutoff.