Coding Agents/Grok Build

Grok Build

by xAI

cliactivesubscription

xAI's agentic coding CLI that drops the Grok models into the terminal with plan mode, parallel subagents, native MCP, and a headless CI mode, plus a one-line Windows PowerShell installer.

Grok Build is xAI's agentic coding CLI: an interactive terminal agent that reads code, runs commands, and edits files the way a developer would, with the Grok models driving it. It entered open beta in May 2026 for SuperGrok and X Premium+ subscribers, and it lands with a notably Windows-friendly setup at a time when many coding agents still treat Windows as an afterthought.

Key capabilities

Plan mode — For a complex task, Grok Build drafts a plan first. You can approve it, comment on individual steps, or rewrite it entirely before a single line of code is touched.

Parallel subagents — For larger jobs it delegates work to specialized subagents that run at the same time, rather than grinding through everything in a single thread.

Native MCP — It speaks the Model Context Protocol out of the box, so the MCP servers you already wired up for other agents — GitHub, Linear, Slack, an internal database — work with no reconfiguration.

Headless / CI mode — The same agent runs non-interactively, so it can sit inside a pipeline instead of only a terminal session.

One-line Windows install — xAI ships a Windows PowerShell installer: irm https://x.ai/cli/install.ps1 | iex.

Model and benchmarks

Under the hood the model scores in the low-70s on SWE-Bench Verified with a 256K-token context window, putting it in the same conversation as Claude Code and Codex CLI. The beta is gated to SuperGrok and X Premium+ subscribers.

Autonomy level

Level 3 (supervised agent): plan mode keeps a human in the loop on the approach before execution, and the agent acts on explicit instruction. Headless / CI mode raises the effective autonomy when run non-interactively in a pipeline.

Strengths

  • First-class Windows support with a one-line PowerShell installer
  • Plan mode gives reviewable, editable plans before any code changes
  • Parallel subagents speed up larger, decomposable tasks
  • Native MCP reuses existing Model Context Protocol servers with no reconfiguration
  • Headless / CI mode lets the same agent run inside a pipeline
  • Competitive coding model: low-70s SWE-Bench Verified, 256K-token context
  • Runs today as a first-class agent inside DevThrottle alongside Claude Code, Codex CLI, and PI

Limitations

  • Not open source
  • Beta access is gated behind a SuperGrok or X Premium+ subscription
  • Locked to the Grok models — no cross-provider model flexibility

Sources

Last verified June 22, 2026