What is claude-architect?

Claude Architect is a Claude Code plugin that lets Claude delegate the actual writing of code to other AI coding tools, such as Codex, OpenCode, Pi, or Pythinker, while Claude itself stays in charge of the specification, the review, and the final judgment call. Each of those outside tools is treated as untrusted: they run in an isolated Git worktree with no memory of the surrounding conversation, and whatever they produce is frozen and checked independently before any of it is allowed to reach the user's actual project files. The plugin enforces three guarantees directly in its code rather than by just asking the AI models to behave a certain way. Every outside tool runs in its own detached worktree with a sanitized environment and a limited list of files it is allowed to touch, so changes outside that scope get rejected. If a tool claims its tests passed, the plugin does not just take its word for it, it reruns the approved verification commands itself in a clean copy of the work and records the real output. And no tool is allowed to approve its own work: review, the human's decision, and the actual merging of code into the project are kept as separate steps, with a person always making the final call. The core workflow moves from a written specification, to one of the outside tools producing code in its isolated worktree, to a frozen candidate version of that code, to independent verification of any tests or checks, to a review step, and finally to a human decision on whether to accept it, reject it, or ask for changes. The project ships as a Claude Code plugin, installed through Claude's plugin marketplace system, and requires Node.js 22 or newer along with at least one of the supported outside coding tools already installed and logged in. It is currently labeled a public beta, and the project explicitly warns against using it without supervision for production, destructive, or security-sensitive work. It is released under the MIT license.