Claude Architect

Verified coding-agent delegation for Claude Code. Claude stays the architect and reviewer — it writes the spec, judges the evidence, and asks you to decide. Implementation is delegated to fresh-context subagent implementers running on the coding CLI you choose — Codex, OpenCode, Pi, or Pythinker — each invocation starting clean with no inherited conversation state, inside an isolated Git worktree. The work comes back as a frozen, hash-anchored candidate that Claude reviews against independent verification evidence before a single byte can reach your checkout.

In practice that means three guarantees the plugin enforces in host code, not in prompts:

• Isolation — every Producer runs in a detached worktree with a sanitized environment, an explicit write allowlist, and OS sandboxing where certified. Out-of-scope changes are rejected at freeze time.
• Evidence over claims — a Producer saying "tests pass" is never accepted; the runtime reruns your authorized verification commands in a clean worktree and records the real output.
• Human acceptance — implementers cannot approve their own work. Review, decision, and hash-gated integration are separate steps, and integration stages the reviewed tree without committing it.

Status

Public beta: Do not use Claude Architect unattended for production, destructive, or security-sensitive work. Review the complete candidate and verification evidence before integration.

The runtime and cross-platform lifecycle are evolving. Producer availability depends on the host OS, CLI version, authentication, requested lane, and proven execution capabilities.

Why it exists

Delegating code generation is easy; establishing which exact bytes were produced, whether they stayed in scope, and whether anyone independent verified them is harder. Claude Architect keeps Claude focused on specification and judgment while treating external coding agents as untrusted Producers. It records a reproducible run, freezes a content-addressed candidate, verifies authorized checks in a clean materialization, and makes the human decision explicit.

Core workflow

flowchart LR
    A[Versioned spec] --> B[Producer in isolated worktree]
    B --> C[Frozen candidate]
    C --> D[Independent verification]
    D --> E[Adversarial review]
    E --> F{Human decision}
    F -->|accept| G[Guarded integration]
    F -->|reject or revise| H[Discard or fresh attempt]

All agent output is an untrusted candidate; implementers cannot approve their own work; only the human accepts.

Installation

Claude Code requires Node.js 22 or newer. Add the marketplace and install the plugin:

claude plugin marketplace add Pythoughts-labs/claude-architect
claude plugin install claude-architect@claude-architect
claude plugin list --json

Restart Claude Code after installing or updating. Install and authenticate at least one supported Producer CLI (codex, opencode, pi, or pythinker); Claude Architect reports unavailable lanes rather than silently substituting another agent.

Quick start

Open Claude Code in a Git repository and name the Producer you want:

/claude-architect:delegate Use Codex to add rate limiting to the public API, run the tests, and show me the independently reviewed candidate before integration.

If no Producer is named, the skill asks you to choose Codex, OpenCode, Pi, or Pythinker. Pi, OpenCode, and Pythinker are harnesses that accept optional model and thinking/variant overrides; model selection within a harness lane is optional and otherwise defers to that CLI's configured default. For non-trivial work it uses the fresh-context review pipeline. Read the exact patch, findings, and verification output before deciding whether to accept.

Available skills, agents, and MCP tools