tulayngmamamo

MCP server enabling bidirectional communication between Claude Code CLI and OpenAI Codex CLI. Allows AI assistants to delegate tasks (research, code review) to each other and receive responses.

Features

• AI-to-AI Communication: Claude and Codex can send messages to each other
• Queue-First Delivery: Fire-and-forget messaging for both Claude and Codex
• Agent Personas: Architect (critical code review) and Oracle (debugging/root cause analysis) personas
• Conversation Tracking: SQLite-backed message history
• memorantado Integration: Sync conversation summaries to knowledge graph
• Unlimited Connections: Multiple Claude Code clients can connect simultaneously

Installation

From npm (recommended)

# Global install
npm install -g tulayngmamamo
tulayngmamamo

# Or run directly with npx
npx tulayngmamamo

From source

git clone https://github.com/alfredosdpiii/tulayngmamamo.git
cd tulayngmamamo
npm install
npm run build
npm start

Server runs at http://127.0.0.1:3790

MCP Configuration

Add to your .mcp.json or MCP client configuration:

Option 1: HTTP Connection (recommended)

Start the server first (tulayngmamamo or npx tulayngmamamo), then configure your MCP client:

{
  "mcpServers": {
    "tulayngmamamo": {
      "type": "http",
      "url": "http://localhost:3790/mcp"
    }
  }
}

Option 2: Via npx (stdio mode)

{
  "mcpServers": {
    "tulayngmamamo": {
      "command": "npx",
      "args": ["tulayngmamamo", "--stdio", "--client-id", "claude"]
    }
  }
}

For Codex, use "codex" as the client id instead of "claude".

MCP Tools

| Tool | Description | |------|-------------| | who_am_i | Returns client identity (claude/codex) | | send_message | Send message to another AI (supports agent parameter) | | delegate_research | Request research on a topic | | request_review | Request code/architecture review | | create_conversation | Start a new conversation | | get_history | Retrieve conversation messages | | share_context | Share files/snippets between AIs | | list_conversations | List active/completed conversations | | close_conversation | Close and optionally summarize a conversation |

Agent Personas

When sending messages to Codex, you can specify an agent persona:

| Agent | Use When | Behavior | |-------|----------|----------| | architect | Architecture questions, code review | Critical analysis, challenges assumptions, examines codebase | | oracle | Debugging, "why" questions | Root cause analysis, strategic reasoning, action steps |

Auto-selected based on message content, or specify explicitly:

send_message(target="codex", content="Review this code", agent="architect")
send_message(target="codex", content="Why is this failing?", agent="oracle")

Architect Persona

The Architect acts as a critical technical partner who:

• Examines the codebase independently before responding
• Agrees when you're right (but explains why)
• Disagrees when they see problems (with specific alternatives)
• Never just rubber-stamps suggestions

Oracle Persona

The Oracle acts as a strategic advisor who:

• Focuses on root cause analysis
• Provides concrete action steps
• Helps understand "why" questions
• Uses strategic reasoning for complex problems

Configuration

| Environment Variable | Default | Description | |---------------------|---------|-------------| | TULAYNGMAMAMO_PORT | 3790 | Server port | | TULAYNGMAMAMO_DB | ~/.tulayngmamamo/tulayngmamamo.sqlite | Database path | | MEMORANTADO_URL | http://127.0.0.1:3789 | memorantado integration URL | | TULAYNGMAMAMO_CLIENT_ID | auto-detect (fallback: claude) | Client identity for stdio mode (claude or codex) | | TULAYNGMAMAMO_CODEX_MODEL | gpt-5.3-codex | Default Codex model | | TULAYNGMAMAMO_CODEX_REASONING_EFFORT | xhigh | Default Codex reasoning effort |

Architecture

Claude Code CLI  <---->  tulayngmamamo  <---->  Codex CLI
       |                      |                    |
       |                      |                    |
    MCP Client           MCP Server          MCP Server
                        (this project)    (codex mcp-server)

Message Flow

1. Client sends message via MCP tool
2. Message is persisted and enqueued in SQLite (message_queue)
3. Queue processor delivers when target client is online
4. Target replies with a new message linked by response_to_id
5. Sender fetches replies via get_response or conversation history

License

MIT