coding-agent-mcp

Local, client-agnostic MCP server for supervising and orchestrating coding agents such as Muse and AGY.

The MCP server is the safe execution, isolation, and verification boundary. OpenAI/ChatGPT, Claude Desktop, custom MCP orchestrators, or any other compatible client act as external supervisors.

Features

Safe Local Execution: Interacts with local agents without exposing arbitrary shell execution.
Git Worktree Isolation: Spawns tasks in isolated Git worktrees by default, protecting current work and enabling safe concurrency.
Client-Agnostic MCP Surface: Standard MCP tools compatible with any client:
- list_agents
- list_repositories
- start_task
- continue_task
- get_task
- get_task_output
- cancel_task
- get_repo_status
- get_diff
- run_verification
Deterministic Verification: Independent test and lint execution profiles returning real exit codes and outputs.
Agent Adapters: Pluggable adapters for Muse and AGY with headless execution, permission management, and session resumption.
Durable Persistence & Audit: SQLite-backed task state, streaming logs, and structured audit logs with crash recovery.

Quick Start

1. Install & Build

npm install
npm run build

2. Configuration

Create ~/.coding-agent-mcp/config.yaml (or copy examples/config.example.yaml):

server:
  data_dir: ~/.coding-agent-mcp
  max_concurrent_tasks: 2

repositories:
  my-repo:
    root: /path/to/my-repo
    writable: true
    default_workspace_strategy: worktree
    verification_profiles:
      test:
        command: ["npm", "test"]
        timeout_seconds: 900
      lint:
        command: ["npm", "run", "lint"]
        timeout_seconds: 300

3. Run Server

node dist/index.js --config /path/to/config.yaml

Running Tests

npm test

Runs the test suite covering unit policies, Git worktree isolation, task lifecycle, and end-to-end MCP tool invocations.

Documentation

License

MIT