Ted Crew MCP Server

Connect Claude Code, Gemini CLI, and Codex CLI as a single MCP server to form an AI mesh network. Each AI can call the others as crew members — self-calls are automatically blocked to prevent infinite loops.

         ┌── ask_codex ──→ Codex ──┐
Claude ──┤                         ├── ask_claude ──→ Claude
         └── ask_gemini ─→ Gemini ─┘
                             │
                             └── ask_codex ──→ Codex

한국어 README

Prerequisites

• Claude Code CLI
• Gemini CLI
• Codex CLI

Installation

Claude Code (~/.claude/settings.json)

{
  "mcpServers": {
    "ted-crew": {
      "command": "npx",
      "args": ["-y", "ted-crew"]
    }
  }
}

Gemini CLI (~/.gemini/settings.json)

{
  "mcpServers": {
    "ted-crew": {
      "command": "npx",
      "args": ["-y", "ted-crew"],
      "env": { "TED_CREW_PROVIDER": "gemini" },
      "trust": true
    }
  }
}

Codex CLI (~/.codex/config.toml)

[mcp_servers.ted-crew]
command = "npx"
args = ["-y", "ted-crew"]
env = { "TED_CREW_PROVIDER" = "codex" }

Cursor (~/.cursor/mcp.json)

{
  "mcpServers": {
    "ted-crew": {
      "command": "npx",
      "args": ["-y", "ted-crew"]
    }
  }
}

Windsurf (~/.codeium/windsurf/mcp_config.json)

{
  "mcpServers": {
    "ted-crew": {
      "command": "npx",
      "args": ["-y", "ted-crew"],
      "disabled": false
    }
  }
}

Cline / VS Code Copilot

{
  "mcpServers": {
    "ted-crew": {
      "command": "npx",
      "args": ["-y", "ted-crew"]
    }
  }
}

Provider Configuration

Set TED_CREW_PROVIDER to identify the caller. The server automatically hides the tool matching the caller to prevent infinite loops.

| Provider | Hidden | Exposed Tools | |----------|--------|---------------| | claude (default) | ask_claude | ask_gemini, ask_codex, jobs | | gemini | ask_gemini | ask_codex, ask_claude, jobs | | codex | ask_codex | ask_gemini, ask_claude, jobs |

IDE-based clients (Cursor, Windsurf, Cline, etc.) can omit TED_CREW_PROVIDER — they'll default to claude.

Use Cases

1. Parallel Code Review

Two AIs review simultaneously from different angles:

ask_gemini(prompt: "Review from an architecture perspective", files: [...], background: true)
ask_codex(prompt: "Review for code quality and bugs", model: "gpt-5.3-codex", reasoning_effort: "high", background: true)
→ wait_job for both → synthesize results

2. Large Codebase Analysis → Implementation

Gemini's 1M token context maps the codebase, then Codex implements:

ask_gemini(prompt: "What files need to change for feature X?", directories: ["./src"], model: "gemini-2.5-pro")
→ feed results to
ask_codex(prompt: "Make the changes based on the analysis", model: "gpt-5.3-codex", reasoning_effort: "high", writable: true)

3. Second Opinion

Get a different perspective on the same problem:

ask_codex(prompt: "Why is this error happening? [error details]", model: "gpt-5.3-codex", reasoning_effort: "xhigh")

4. Documentation Pipeline

Draft with writing-focused Gemini → Claude refines with project context:

ask_gemini(prompt: "Read this code and write a README draft", files: ["./src/index.ts"])
→ Claude edits the draft to fit the project

5. Research → Implementation

Outsource research to preserve Claude's context window:

ask_gemini(prompt: "Summarize React Query v5 staleTime vs gcTime differences", model: "gemini-2.5-pro")
→ Claude applies the findings to actual code

AI Strengths at a Glance

| AI | Strengths | |----|-----------| | Gemini | Large context analysis (1M tokens), writing, research | | Codex | Code execution/editing, build/test, tunable reasoning_effort | | Claude | Orchestrator, conversation context, final judgment |

Tools

ask_gemini

Delegates tasks to Gemini CLI. Best for 1M token context, writing, and research.

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | prompt | string | Yes | Prompt text | | model | string | | Model override | | files | string[] | | File paths to inject as context | | directories | string[] | | Directories for Gemini to scan directly | | output_file | string | | File path to save the response | | working_directory | string | | Working directory | | background | boolean | | Run in background | | approval_mode | string | | yolo / auto_edit / plan |

ask_codex

Delegates tasks to Codex CLI. Directly control behavior with model and reasoning_effort.

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | prompt | string | Yes | Prompt text | | model | string | | Model override (default: Codex CLI default) | | reasoning_effort | string | | minimal / low / medium / high / xhigh | | files | string[] | | Context file paths | | output_file | string | | File path to save the response | | working_directory | string | | Working directory | | background | boolean | | Run in background | | writable | boolean | | Allow file modifications | | timeout_ms | number | | Foreground timeout in ms (default: 300,000) |

ask_claude

Delegates tasks to Claude Code CLI. Best for code generation, debugging, and refactoring.

| Parameter | Type | Required | Description | |-----------|------|----------|-------------| | prompt | string | Yes | Prompt text | | model | string | | claude-opus-4-6 / claude-sonnet-4-6 / claude-haiku-4-5 | | files | string[] | | Context file paths | | output_file | string | | File path to save the response | | working_directory | string | | Working directory | | background | boolean | | Run in background | | allowed_tools | string[] | | Allowed tools (Read, Write, Edit, Bash, etc.) |

Job Management

Manage background jobs.

| Tool | Description | |------|-------------| | wait_job | Wait for completion, return full stdout/stderr | | check_job | Non-blocking status check, 500-char stdout preview | | kill_job | Force terminate | | list_jobs | List jobs (filter: active/completed/failed/all) |

Environment Variables

| Variable | Default | Description | |----------|---------|-------------| | TED_CREW_PROVIDER | claude | Caller identity (claude / gemini / codex) | | TED_CREW_MAX_STDOUT | 10485760 | stdout collection limit (10MB) | | TED_CREW_TIMEOUT | 300000 | Default foreground timeout (5 min) |

Auto-Save Responses

Responses longer than 500 characters are automatically saved to .aidocs/ted-crew/{provider}-{date}-{time}.md and only a summary is returned. Specify working_directory to save under that project's directory.

Install Skills

Install routing guide skills for each AI client:

npx ted-crew install-skills

Or from source:

npm run install-skills

Install locations:

| AI | Path | |----|------| | Claude Code | ~/.claude/skills/ted-crew/ | | Gemini CLI | ~/.gemini/skills/ted-crew/ | | Codex CLI | ~/.codex/skills/ted-crew/ |

Project Structure

src/
├── index.ts                 # MCP server entry, tool filtering
├── lib/
│   ├── constants.ts         # Shared constants
│   ├── exchange.ts          # Response post-processing (file save)
│   ├── parser.ts            # CLI output parsing, error detection
│   ├── prompt-builder.ts    # Prompt construction
│   ├── spawner.ts           # CLI process spawning
│   └── types.ts             # TypeScript type definitions
└── tools/
    ├── ask-claude.ts        # Claude Code CLI caller
    ├── ask-codex.ts         # Codex CLI caller
    ├── ask-gemini.ts        # Gemini CLI caller
    └── jobs.ts              # Background job manager

Build

npm install
npm run build    # esbuild → dist/server.cjs
npm run dev      # watch mode

License

MIT