SpaceXAI Open-Sources Grok Build: The Rust Agent Harness, TUI, and Tool Layer Behind Its Coding CLI

SpaceXAI put the code behind its grok CLI out into the open today, dropping the whole agent harness, terminal UI, CLI shell, and developer tooling onto GitHub under an Apache 2.0 license. Grok Build itself isn't brand new — it launched as an early beta back on May 25, 2026 — but what changes now is that anyone can read, fork, and rebuild the thing that reads your codebase, edits files, runs shell commands, searches the web, and keeps track of long-running tasks while it does it, all from a full-screen, mouse-interactive terminal window.

A harness, if you haven't dealt with one before, is basically the plumbing around a language model: it gathers context, sends the prompt, parses whatever comes back, and figures out which tools to call. SpaceXAI splits Grok Build's published code into four areas — the agent loop, the tools, the terminal UI, and an extension system covering skills, plugins, hooks, MCP servers, and subagents — and maps them to named Rust crates. Want to understand the loop first? Start with xai-grok-shell, then move to xai-grok-tools. One thing worth flagging for anyone poking around the repo: the root Cargo.toml is auto-generated, and the README explicitly says to leave it alone.

The more interesting story here might be the local-first angle. Compile Grok Build yourself, point it at your own inference endpoint, and configure the whole thing through a config.toml file that lets you register a model, its base URL, and an API key under whatever display name you like. A grok inspect command then reports back everything the harness found in your working directory — config sources, instructions, skills, plugins, hooks, MCP servers, the works. For teams that don't want a coding agent phoning home to api.x.ai, that's a real option, not just a talking point.

But open-sourcing the code doesn't mean SpaceXAI wants your patches. The project accepts forks and modifications freely under Apache 2.0, and you can wire it up to any model through config.toml, yet external pull requests aren't accepted. That puts Grok Build in an odd middle ground next to comparable projects: OpenAI's Codex CLI is also Apache 2.0 and forkable, but keeps an open PR queue for OpenAI models; OpenCode is MIT-licensed, community-run, and supports 75+ providers; Anthropic's Claude Code, by contrast, stays proprietary and closed to forking entirely, tied to Anthropic's own models.

What people will actually do with this, according to SpaceXAI's own framing, breaks down into a handful of practical paths — auditing the tools crate before trusting an agent with shell access in a regulated repo, forking the harness for an internal build knowing upstream won't take it back, running it air-gapped by pointing base_url at an internal endpoint, or piping headless output into a CI pipeline step. None of that requires SpaceXAI's blessing going forward. That's the whole point of Apache 2.0.