Grok Build Is Free and Open Source. Here's What xAI Left Out of the Announcement

xAI open-sourced Grok Build, its terminal coding agent, in mid-July 2026. The Rust source is public on GitHub under `xai-org/grok-build`, it picked up roughly 4,500 stars in its first few weeks, and it runs on Grok 4.5 with a genuinely polished full-screen terminal interface. The framing was generous: free, open, first-party from a major lab.

All of that is true. It's also not the whole picture, and the gaps are worth knowing before you install it.

What Grok Build actually does

Grok Build isn't a chat window wearing a code editor's clothes. It's an agent harness, a loop that reads your codebase, plans a change, edits files, runs shell commands, and checks its own work. Three ways to run it: an interactive terminal UI with a native subagent view and Plan Mode, a headless mode for CI pipelines, and integration through the Agent Client Protocol so editors like Zed can use it as their backing agent.

Installation is a one-line script:

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

It opens your browser to authenticate against your xAI account on first launch, so there's no manual key-pasting for the default path. If piping a script straight into your shell makes you uneasy, that's a reasonable instinct for any installer, not something specific to xAI. Download it, read it, then run it.

The real cost, worked out

The tool itself is free. What you pay for is Grok 4.5 through xAI's API, priced at roughly $2 per million input tokens and $6 per million output tokens. A single moderate agent turn, say 30K tokens of context in and 5K tokens of edits out, comes to:

(30 × $2 / 1000) + (5 × $6 / 1000) ≈ $0.06 + $0.03 = $0.09 per task

Under a dime for one editing pass. A real feature build with tests and a few rounds of iteration will run several of those turns, but the per-token math puts Grok 4.5 in the cheap-and-fast tier, comparable to Gemini's Flash line and noticeably cheaper than routing the same work through Claude Opus-class models.

Open source, but not open development

Here's the part that got less airtime than the "open source" headline deserved. The public repo is a periodic export from xAI's internal monorepo, not a live-developed project. There's no commit history, and the project explicitly does not accept external pull requests. You can read it, fork it, and modify your own copy under Apache 2.0, but you can't send anything upstream. It's open source in the license sense, closed in the development sense.

Simon Willison dug into the tools crate after launch and found something the repo itself discloses in `THIRD_PARTY_NOTICES.md`: a meaningful chunk of the file-editing toolchain is ported from other agents. The patch-apply, grep, and directory-listing handlers trace back to OpenAI's Codex CLI. The bash, edit, glob, read, and write handlers trace back to OpenCode. Willison's read was that the ports are license-compliant, properly credited under Apache §4(b), and pragmatic rather than sneaky. xAI built the TUI, the agent runtime, and the workspace/checkpoint layer from scratch; it reused battle-tested tool handlers instead of reinventing file editing. Whether that counts as smart reuse or as leaning on someone else's hard work is a matter of taste, and both readings showed up in the discussion after launch.

One more detail people enjoyed poking at: the subagent system prompt reportedly instructs the model not to reveal its own contents. Small, a little self-conscious, and telling.

How it stacks up

If you're already inside the xAI ecosystem and want an official, fast, cheap agent with a well-built terminal UI, Grok Build is a solid default. If model choice matters more to you than any single vendor's speed, OpenCode is still the pick, if only because it's model-agnostic by design, and a little ironically, Grok Build borrows several of its own tool implementations from OpenCode anyway. [Kilo](https://questloops.com/tools/kilo) takes a similar model-agnostic approach with a real free tier across 500+ models, worth a look if vendor lock-in is your main objection here. If an active outside-contributor community matters, Codex and OpenCode both beat a monorepo export with the PR door closed.

Honest limitations

• You're locked to Grok 4.5 through xAI's API. There's no "swap to Claude or Gemini" option baked in the way OpenCode offers.
• Windows builds from source are untested by the maintainers; prebuilt Windows binaries exist, but macOS and Linux are the first-class targets.
• Building from source requires DotSlash on your PATH to resolve a vendored `protoc` dependency, an extra step most Rust projects don't ask for.
• The tool is free. The model isn't. Multi-turn feature builds and web-search-heavy tasks add up fast at $2/$6 per million tokens.

FAQ

**Is Grok Build actually free?** The tool is free and open source under Apache 2.0. You pay xAI directly for Grok 4.5 API usage, billed through your xAI account rather than a separate key.

**Can I contribute to it?** No. It's a periodic export from an internal monorepo with no commit history, and xAI has said explicitly that external pull requests aren't accepted.

**Can I point it at Claude or GPT instead of Grok?** No. It's wired to xAI's models. If you want model choice, use OpenCode instead.

**Is it just a Codex fork?** No, though it borrows tool code from Codex and OpenCode, disclosed and license-compliant. The agent loop, terminal UI, and checkpoint system are first-party xAI work.

The bigger picture: every major lab now ships its own first-party terminal coding agent, and Grok Build is a credible one. Go in knowing the $0.09-a-task math and the closed-contribution model going in, and there's nothing here that should surprise you later.