使用说明

Co-Task MCP

An MCP server that spawns parallel coding agents to work on repositories autonomously. Delegate tasks to Copilot CLI, Codex CLI, or Claude Code and get PRs back.

Features

🚀 Parallel Execution - Run multiple coding tasks simultaneously
🤖 Multi-Agent Support - Use Copilot CLI, Codex CLI, or Claude Code
📦 Automatic Git Workflow - Clones repos, creates branches, opens PRs
🔔 Native Notifications - macOS notifications when tasks complete
🧹 Auto-Cleanup - Removes workspaces after PRs merge
🔒 Session Isolation - Tasks are scoped to the session that created them
⏹️ Graceful Shutdown - Tasks cancel when parent session closes

Installation

Prerequisites

Python 3.10+
uv package manager
Git and GitHub CLI (gh)
At least one coding CLI:
- co - GitHub Copilot CLI
- codex - OpenAI Codex CLI
- claude - Claude Code CLI

Install

git clone https://github.com/crmitchelmore/co-task-mcp
cd co-task-mcp
uv sync

Configure MCP Client

Add to your MCP client configuration (e.g., ~/.copilot/mcp-config.json):

{
  "mcpServers": {
    "multitasker": {
      "command": "uv",
      "args": ["run", "--directory", "/path/to/co-task-mcp", "python", "-m", "multitasker_mcp.server"],
      "type": "local",
      "tools": [
        "multitasker_run",
        "multitasker_status", 
        "multitasker_logs",
        "multitasker_list",
        "multitasker_cancel",
        "multitasker_cleanup",
        "multitasker_answer",
        "multitasker_wait"
      ]
    }
  }
}

Quick Start

Start a Single Task

Use multitasker to add input validation to https://github.com/user/repo

Start Parallel Tasks

Use multitasker to run these 3 tasks on https://github.com/user/repo:
1. Add TypeScript types for the API layer
2. Create React components for the dashboard  
3. Set up Zustand state management

Monitor Progress

The agent will automatically poll multitasker_list every 60 seconds until all tasks complete. You'll get macOS notifications as tasks finish.

Tools

Tool	Description
`multitasker_run`	Start a background coding task
`multitasker_list`	List all tasks from this session
`multitasker_status`	Get status of a single task
`multitasker_logs`	View task output logs
`multitasker_cancel`	Cancel a running task
`multitasker_cleanup`	Clean up workspace after PR merged
`multitasker_wait`	Wait for tasks with progress updates
`multitasker_answer`	Answer questions from tasks (human-in-the-loop)

Resources

The MCP server exposes these resources:

URI	Description
`multitasker://runs`	List all runs
`multitasker://runs/{run_id}/logs`	Full logs for a run
`multitasker://runs/{run_id}/state`	JSON state for a run

Configuration

Create ~/.multitasker/config.json to customize agent commands:

{
  "commands": {
    "copilot": {"cmd": "co", "args": [], "model_flag": "--model", "prompt_flag": "-p"},
    "codex": {"cmd": "codex", "args": [], "prompt_flag": null},
    "claude": {"cmd": "claude", "args": ["-p"], "prompt_flag": null}
  },
  "copilot_models": {
    "opus": ["claude-sonnet-4"],
    "gpt": ["gpt-4.1"],
    "gemini": ["gemini-2.5-pro"]
  }
}

How It Works

┌─────────────┐     ┌─────────────┐     ┌─────────────┐
│  MCP Client │────▶│  MCP Server │────▶│   Worker    │
│  (Copilot)  │     │  (FastMCP)  │     │  (Process)  │
└─────────────┘     └─────────────┘     └─────────────┘
                           │                   │
                           ▼                   ▼
                    ┌─────────────┐     ┌─────────────┐
                    │    State    │     │  Coding CLI │
                    │ (~/.multi*) │     │ (co/cx/cc)  │
                    └─────────────┘     └─────────────┘

Task Submission - Call multitasker_run with repo URL and prompt
Clone & Branch - Worker clones repo and creates feature branch
Agent Execution - Spawns coding CLI with your prompt
PR Creation - Creates PR via gh when complete
Cleanup - Workspace removed after PR merges

Documentation

See the docs folder for detailed documentation:

Requirements

Python 3.10+
git and gh CLI tools
One of: co (Copilot CLI), codex (Codex CLI), claude (Claude Code)

License

MIT