All-Agents-MCP

⚠️ Archived — This project is no longer actively maintained.

As the AI agent CLI ecosystem has evolved, direct CLI invocation + Skills-based integration has become the more practical and mainstream approach over wrapping agent CLIs behind an MCP server.

The core functionality of this project has been merged into hardened-claude-code as Skills. The same multi-agent orchestration can now be achieved more concisely and maintainably through CLAUDE.md configuration and Skills.

Migration guide:

• Tools like ask_agent, ask_all, delegate_task → replaced by Skills in hardened-claude-code (/with, /web-fetch, etc.)

• No MCP server setup required — Skills invoke CLIs directly

• Environment variable model config → consolidated into declarative CLAUDE.md settings

Original README below for reference.

An MCP server that orchestrates multiple AI CLI agents — Claude Code, Codex, Gemini CLI, and Copilot CLI — through a unified interface. Delegate tasks, run cross-agent comparisons, and leverage each agent's strengths from any MCP-compatible host.

Safe by Design — No OAuth Token Hijacking

Some multi-agent tools work by extracting OAuth tokens from other AI services and calling their APIs directly. This approach (used by projects like OpenCode) violates the terms of service of those platforms and can result in account suspension or permanent bans.

all-agents-mcp takes a fundamentally different approach. It invokes each agent's official CLI binary (claude, codex, gemini, copilot) as a child process — exactly the same way a human would use them in a terminal. No tokens are extracted, no APIs are called behind the scenes, and no authentication is bypassed.

Each CLI agent manages its own authentication, billing, and rate limits. all-agents-mcp is simply a process orchestrator — it doesn't touch your credentials.

Features

• Single-agent queries — Ask a specific agent with ask_agent

• Multi-agent comparison — Run the same prompt across all agents in parallel with ask_all

• Task delegation — Auto-analyze complexity and route to one or multiple agents

• Cross-model verification — Verify answers by running one agent with different models

• Specialized tools — Code review, debugging, explanation, test generation, refactoring

• Recursive call prevention — Automatically excludes the calling agent to avoid infinite loops

• Session history — All interactions are recorded and queryable via MCP resources

• Environment-based model config — Override models at runtime via AA_MCP_* environment variables

Prerequisites

• Node.js 22+

• At least one of the following CLI agents installed and authenticated:

Installation

Claude Code

As a Plugin (recommended)

# 1. Add marketplace (includes all Dokkabei97 plugins)
/plugin marketplace add Dokkabei97/claude-plugins

# 2. Install plugin
/plugin install all-agents-mcp

You can also add the plugin repository directly: /plugin marketplace add Dokkabei97/all-agents-mcp

This installs all-agents-mcp as a Claude Code plugin, giving you:

• 8 skills (/all-agents-mcp:ask, /all-agents-mcp:ask-all, /all-agents-mcp:delegate, /all-agents-mcp:review, /all-agents-mcp:debug, /all-agents-mcp:agents, /all-agents-mcp:aa-models, /all-agents-mcp:aa-fetch)

• Automatic MCP server connection via npx

• Session start health checks

As MCP Server

claude mcp add all-agents-mcp -- npx -y all-agents-mcp

Or manually add to ~/.claude.json:

{
  "mcpServers": {
    "all-agents-mcp": {
      "command": "npx",
      "args": ["-y", "all-agents-mcp"]
    }
  }
}

Codex CLI

codex mcp add all-agents-mcp -- npx -y all-agents-mcp

Or manually add to ~/.codex/config.toml:

[mcp_servers.all-agents-mcp]
command = "npx"
args = ["-y", "all-agents-mcp"]

[mcp_servers.all-agents-mcp.env]
AA_MCP_LOG_LEVEL = "warn"

Gemini CLI

gemini mcp add all-agents-mcp npx -y all-agents-mcp

Or manually add to ~/.gemini/settings.json:

{
  "mcpServers": {
    "all-agents-mcp": {
      "command": "npx",
      "args": ["-y", "all-agents-mcp"]
    }
  }
}

Copilot CLI

Add to ~/.copilot/mcp-config.json:

{
  "servers": [
    {
      "name": "all-agents-mcp",
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "all-agents-mcp"]
    }
  ]
}

From Source

git clone https://github.com/Dokkabei97/all-agents-mcp.git
cd all-agents-mcp
npm install
npm run build

# Claude Code
claude mcp add all-agents-mcp -- node /path/to/all-agents-mcp/dist/index.js

# Codex
codex mcp add all-agents-mcp -- node /path/to/all-agents-mcp/dist/index.js

# Gemini CLI
gemini mcp add all-agents-mcp node /path/to/all-agents-mcp/dist/index.js

Plugin Skills

When installed as a Claude Code plugin, the following skills are available:

Skills

Tools (14)

Core Tools

Verification

Specialized Tools

Web Tools

Info Tools

Resources (3)

Usage Examples

Once registered, use natural language in Claude Code:

# Ask a specific agent
"Ask Codex to implement a binary search function"

# Compare across agents
"Ask all agents how to optimize this SQL query"

# Delegate a task
"Delegate writing unit tests for src/utils/ to Gemini"

# Collaborate
"Collaborate with Copilot to review this architecture"

# Cross-model verification
"Verify with Copilot using gpt-5.2-codex, claude-sonnet-4.5, and gemini-3-pro-preview"

# Code review
"Ask Codex to review this code for security vulnerabilities"

# Debug
"Debug this TypeError with Gemini"

# Generate tests
"Generate vitest tests for this function using Codex"

# Check status
"Show me all available agents and their health"

Model Configuration

Models are configured via environment variables with the AA_MCP_ prefix. Set them in your MCP client config to override defaults at runtime:

{
  "mcpServers": {
    "all-agents-mcp": {
      "command": "npx",
      "args": ["-y", "all-agents-mcp"],
      "env": {
        "AA_MCP_CLAUDE_DEFAULT": "claude-sonnet-4.5",
        "AA_MCP_CODEX_ANALYSIS_LEVEL": "medium",
        "AA_MCP_GEMINI_MODELS": "gemini-2.5-pro,gemini-2.5-flash"
      }
    }
  }
}

Available Environment Variables

Recursive Call Prevention

When all-agents-mcp runs inside an agent (e.g., Claude Code calls all-agents-mcp), it automatically detects the caller and excludes it from the available agent list. This prevents infinite recursive loops.

Detection methods (in priority order):

1. CLI argument: --caller=claude

2. Environment variables: CLAUDECODE, CODEX_SANDBOX_TYPE, GEMINI_CLI, COPILOT_CLI

3. process.env._ fallback

Development Methodology

This project is built following AIDE (Agent-Informed Development Engineering) v1.0 — a software development methodology designed for the agentic era. Key principles applied include context budget constraints, locality of behavior, functional core architecture, test-as-specification, and deterministic guardrails.

Test Status

All agents have been tested and verified.

Architecture

all-agents-mcp/
├── src/
│   ├── agents/          # Agent abstraction layer
│   │   ├── types.ts     #   IAgent interface, AgentResponse, HealthStatus
│   │   ├── base-agent.ts#   Abstract base with spawn logic
│   │   ├── claude-agent.ts
│   │   ├── codex-agent.ts
│   │   ├── gemini-agent.ts
│   │   ├── copilot-agent.ts
│   │   └── registry.ts  #   Detection, registration, recursion guard
│   │
│   ├── tools/           # 14 MCP tool definitions
│   ├── orchestrator/    # Parallel execution, complexity analysis, verification
│   ├── session/         # File-based session storage
│   ├── resources/       # 3 MCP resource definitions
│   ├── config/          # Env var config loader (AA_MCP_* overrides) + Zod schema
│   ├── utils/           # Logger (stderr), CLI detection
│   ├── server.ts        # McpServer factory
│   └── index.ts         # Entry point (stdio transport)

Development

npm install          # Install dependencies
npm run build        # Compile TypeScript
npm run dev          # Watch mode
npm run lint         # Biome lint + format
npm test             # Run tests

Debug logging

AA_MCP_LOG_LEVEL=debug node dist/index.js

Log levels: debug, info (default), warn, error. All logs go to stderr to avoid interfering with MCP stdio transport.

License

MIT