aming-claw

Aming Claw

Your AI agent and you, sharing the same dashboard.

Open source workspace for AI-coded development. See what AI is touching in real-time, audit every change, review proposals before they land in your codebase. Multi-language (Python/TypeScript), MCP-native, local-first.

AI-Agent Review Control Plane

Aming Claw is a graph-governed control plane for AI-agent code review. The claim is narrow: not total program correctness, not human-free review, and not free first-build cost on huge monorepos. The V1 loop is a reusable project fact layer: commit-bound graph state, typed AI proposals, review gates, durable repair rules, and reconcile after code or rules change.

flowchart TD
  Human["Human reviewer"] --> UI["Dashboard"]
  Agent["AI agent"] --> MCP["MCP tools"]

  UI --> Gov["Governance control plane"]
  MCP --> Gov

  Gov --> Backlog["Backlog<br/>intent + work ledger"]
  Gov --> Review["Review Queue<br/>accept / reject proposals"]
  Gov --> Graph["Commit-bound graph<br/>files, functions, tests, docs, config"]

  Agent --> Proposal["Typed AI proposals<br/>semantic / graph / config / backlog"]
  Proposal --> Review

  Code["Committed code"] --> Reconcile["Reconcile<br/>derive review state"]
  Rules["Hints + config + rules"] --> Reconcile
  Review --> Events["Accepted semantic events"]
  Events --> Reconcile

  Reconcile --> Graph
  Reconcile --> Semantics["Semantic projection<br/>current / stale / missing"]

  Graph --> Context["Scoped review context"]
  Semantics --> Context
  Backlog --> Context

The important boundary: AI proposals are not trusted state. Committed code, source-controlled hints/config/rules, and accepted semantic events are source; graph snapshots, semantic projections, review queues, and test/doc/config bindings are derived state.

Contents

• Get Started
• HN Demo
• Review-Scope Challenge
• Key Terms
• Architecture Contracts
• Demos
• Dashboard Surfaces
• Use Cases
• Observer-Led Manual Fix
• V1 MVP Snapshot
• Install Details
• Runtime Boundaries
• Plugin Components
• Workflow
• CLI
• Packaging
• Governance Contract
• Deep Dive
• Status & Contributing
• Documentation
• License

Get Started

Start with the AI host you use. The longer requirements, raw installer commands, and troubleshooting notes are in Install Details.

Codex

Ask Codex for the complete one-shot path. This installs the plugin/runtime, starts local governance, and opens the dashboard:

One-shot install and open dashboard for Aming Claw from https://github.com/amingclawdev/aming-claw

If you only want to refresh the local plugin package, use the install-only command instead. Install-only does not start the long-running governance service and does not open the dashboard:

aming-claw plugin install https://github.com/amingclawdev/aming-claw

Reload Codex or open a new Codex session after install. First-run startup and verification steps are in Install Details.

Claude Code

Paste this once — Claude bootstraps the plugin end-to-end:

Install aming-claw end-to-end from https://github.com/amingclawdev/aming-claw:
1. Run `/plugin marketplace add https://github.com/amingclawdev/aming-claw`
2. Run `/plugin install aming-claw@aming-claw-local`
3. pip install -e the marketplace clone at
   ~/.claude/plugins/marketplaces/aming-claw-local
   (Windows: %USERPROFILE%\.claude\plugins\marketplaces\aming-claw-local)
4. Start `aming-claw start` in a background terminal
5. Run `aming-claw open` to launch the dashboard
6. Remind me to reload Claude Code so the plugin's MCP tools and skills load

Steps 1–2 only copy the skill files; pip install adds the Python runtime, aming-claw start boots the local governance service, and aming-claw open launches the dashboard. After reloading, phrases like "install and start aming-claw" or "one-shot install" trigger the launcher skill's one-shot mode for re-bootstrap in future sessions. First-run troubleshooting and raw installer scripts are in Install Details.

After installation: restart or open a new host session

The Claude Code or Codex session you installed or updated Aming Claw in won't load newly added plugin skills or MCP tools. The host reads plugin paths at startup, so changes apply only after a real host reload/restart or a fresh session that reindexes plugins.

Works immediately (no new session needed):