Grok Build: the terminal coding agent, its plans and its open-source code

What Grok Build is

Grok Build is the coding agent from SpaceXSI (formerly xAI). It runs in your terminal as a fullscreen interface with mouse support. SpaceXSI says it works with any codebase and any language. It can read a project, edit files, run shell commands, search the web and manage long-running tasks. Grok 4.7 now powers it.

It entered beta on May 14, 2026, according to the release notes, and the formal launch post followed on May 25. Since then SpaceXSI has added an Agent Dashboard (June 15), a Plugin Marketplace (June 11), Workflows (July 23), Build on web and mobile (August 19) and Memory (September 16). For setup steps, see how to install Grok Build.

Plans and price

The official Grok Build page says it is “available to try for free.” The X Subscription pricing page lists Build in the Plus plan ($8 per month), where it shares one usage pool with Grok chat, voice, Imagine and API access. Higher tiers (Premium $30, Super $100, Ultra $200) give more of that shared pool.

• Free: try it with limited usage. Grok 4.7 Fast is not included.
• Plus and up: Build is part of the shared weekly pool.
• API key: you can also run it against pay-as-you-go API credit, for example in CI.

Key features

• Plan mode: no files change until you approve a plan. You can approve it, comment on single steps or rewrite it. Approved changes appear as diffs.
• Subagents: several agents run in parallel, each with its own context window. Git worktrees are supported. Launch coverage said up to eight agents could run at once, but the current official page gives no maximum.
• Skills and plugins: skills run automatically when a task matches, or can be called by name. The /skillify command turns a session into a skill. Plugins bundle skills, agents, hooks and MCP servers.
• MCP: connects to services such as Linear, Sentry, Postgres, Grafana and browsers.
• AGENTS.md: sets coding conventions per directory. When a task is ambiguous, Build asks multiple-choice questions.
• Headless mode: runs without the interface for scripts and CI/CD pipelines.
• Editor integration: through the Agent Client Protocol (ACP).

The official page also lists hooks, memory, code search, git integration, web search, code review, sandboxing, background tasks and themes.

Open source since July 15, 2026

SpaceXSI published the full source on July 15, 2026, at github.com/xai-org/grok-build. Some outlets say July 16, which is the press date. First-party code is under the Apache 2.0 license, while vendored third-party code keeps its own licenses. When we checked, the repo had about 27.4k stars and 5.2k forks.

Simon Willison counted about 844,530 lines of Rust, released as a single commit. The repo includes the system prompt and the subagent prompt. Some tools were ported from OpenAI’s Codex and from OpenCode. External contributions are not accepted. The models themselves remain closed; see open-source Grok models for what has been released.

The July 2026 upload incident

In mid-July 2026, users found that running grok in a directory could upload that whole directory to the company’s Google Cloud storage. One user who ran it in a home directory said SSH keys, a password-manager database and personal files were uploaded. Wikipedia dates the reports to July 14 and says the behavior was disabled the same day.

What changed afterward:

• Musk said previously uploaded data would be deleted, and the upload feature was turned off.
• SpaceXSI says default data retention has been off for all Grok Build users since July 12, and retained coding data is being deleted.
• The source went public to “rebuild trust,” and Build can now run fully locally.
• The open code still contains the upload code, but it is disabled and returns a hard-coded “unavailable” error, according to Simon Willison.

grok-build-0.1 and benchmarks

SpaceXSI also sells a coding model on its API called grok-build-0.1. Early access began May 19, 2026. It has a 256k-token context window and costs $1.00 per million input tokens, $0.20 cached and $2.00 output. Prompts over 200k tokens cost double. Calls to the retired grok-code-fast-1 now route to it.

Benchmark claims need care:

• The 70.8% SWE-bench Verified figure tied to Grok Build at launch actually belongs to its predecessor, grok-code-fast-1, measured on xAI’s own harness in August 2025. SpaceXSI has published no SWE-bench score for grok-build-0.1.
• Independent figures for grok-build-0.1, as reported by Agentriot: 71.4% on Vals AI SWE-bench, 50.6% on Terminal-Bench 2.0 and 88.9% on PinchBench.
• Grok 4.7 running inside Grok Build scored 56 on the Artificial Analysis Coding Agent Index, fourth among native harnesses (September 2026).

Prefer to code in an editor? Grok models also run in Cursor.

Questions people ask