waha-langchain-mcp

MCP server that exposes a local WAHA (WhatsApp HTTP API) instance as tools for LangChain ReAct agents and deep agents.

flowchart LR
    subgraph Agent["LangChain process"]
        A["ReAct agent"]
        D["Deep agent"]
        C["MCP client"]
        A --> C
        D --> C
    end
    subgraph Server["waha-langchain-mcp (stdio)"]
        T["MCP tools"]
        W["WahaClient"]
        T --> W
    end
    C -->|"JSON-RPC / stdio"| T
    W -->|"HTTP + X-Api-Key"| WAHA["WAHA :3000"]
    WAHA <--> WhatsApp

Full diagram: docs/architecture/waha-mcp.mmd. Design decision: docs/decisions/0001-mcp-server-for-waha.md.

Tools

Sessions: waha_status, waha_list_sessions, waha_get_session, waha_get_me, waha_start_session, waha_stop_session, waha_create_session.

Chats and messages: waha_list_chats, waha_get_chat, waha_get_chat_messages, waha_mark_chat_read, waha_set_typing, waha_take_screenshot.

Sending: waha_send_text, waha_send_image, waha_send_file.

Contacts: waha_check_number_exists, waha_get_profile_picture.

Inbound messages are not pushed over MCP. For that, point a WAHA webhook at your application; MCP is request/response only.

Related MCP server: WA MCP

Setup

cd /home/user/Projects/aegra-content/waha-langchain-mcp
uv venv
uv pip install -e ".[examples]"
cp .env.example .env

Set WAHA_API_KEY in .env.

The running container was started without --env-file, so WAHA generated a random key at startup. Read it with:

docker logs waha 2>&1 | grep -A2 "WAHA_API_KEY"

To stop rotating keys on restart, run the container with your env file:

docker run --rm -p 3000:3000 --name waha \
  --env-file /home/user/Projects/aegra-content/services-waha/.env \
  -v /home/user/Projects/aegra-content/services-waha/sessions:/app/.sessions \
  devlikeapro/waha

The dashboard username/password (WAHA_DASHBOARD_*) are basic-auth credentials for the UI; the MCP server uses WAHA_API_KEY / X-Api-Key.

Verify the server and the live WAHA connection:

uv run python examples/list_tools.py

Use with LangChain

The same MCP tool list feeds both agent types.

ReAct agent:

uv run python examples/react_agent.py "Show my 5 most recent chats"

Deep agent:

uv run python examples/deep_agent.py "Summarize my unread chats"

mcp is pinned below 2.0 because langchain-mcp-adapters still requires the 1.x SDK.

Both examples spawn python -m waha_mcp over stdio via langchain_mcp_adapters.client.MultiServerMCPClient. The API key is passed in the child process environment, never to the model.

Minimal wiring:

from langchain_mcp_adapters.client import MultiServerMCPClient

client = MultiServerMCPClient({
    "waha": {
        "command": "python",
        "args": ["-m", "waha_mcp"],
        "transport": "stdio",
        "env": {"WAHA_BASE_URL": "http://localhost:3000", "WAHA_API_KEY": "..."},
    }
})
tools = await client.get_tools()

Then create_agent(model, tools) or create_deep_agent(model=model, tools=tools).

Use with other MCP hosts

Claude Desktop / Cursor style config:

{
  "mcpServers": {
    "waha": {
      "command": "uv",
      "args": ["run", "--directory", "/home/user/Projects/aegra-content/waha-langchain-mcp", "waha-mcp"],
      "env": {
        "WAHA_BASE_URL": "http://localhost:3000",
        "WAHA_API_KEY": "<your-key>"
      }
    }
  }
}

Streamable HTTP (optional):

WAHA_MCP_TRANSPORT=streamable-http python -m waha_mcp

Print the tool catalog without starting the server:

python -m waha_mcp --pretty

Security

• Keep WAHA_API_KEY in the environment or .env; it is never exposed as a tool.

• Sending messages is a real side effect. Keep explicit user intent in the agent system prompt and add human approval where the host supports it.

• The stdio server has the same filesystem access as the agent process because waha_send_file / waha_send_image accept local paths.

License

MIT. See LICENSE.