mcp-telegram

An MCP server that lets an LLM send messages, photos and documents to Telegram through a bot.

Point any MCP-capable client (Claude Desktop, Claude Code, or your own agent) at this server and the model gains four tools: get_me, send_message, send_photo and send_document. The model decides what to say; the server handles how it reaches Telegram.

Why the Bot API

This server talks to the Telegram Bot API. That means you create a bot, the bot sends messages, and users receive them.

• No phone-number login, no api_id/api_hash, no session files.

• The bot can only message chats it already shares with a user (the user must start the bot first) or channels/groups it belongs to.

If you instead need to send messages as your own user account to arbitrary people, that requires the MTProto client API (Telethon/Pyrogram) and account authentication — a deliberately different, heavier tool that is out of scope here.

Related MCP server: mcp-telegram

Architecture

Three small layers, each with one responsibility:

src/mcp_telegram/
├── config.py     # load & validate settings from the environment (fail-fast)
├── telegram.py   # thin async client over the Bot API (httpx, no heavy SDK)
├── server.py     # FastMCP server: tool definitions + shared client lifespan
└── __main__.py   # entry point (stdio transport)

The MCP layer never builds HTTP requests and the HTTP layer never knows about MCP, so each can be tested and changed on its own. A single httpx.AsyncClient is created once per server run via the FastMCP lifespan and reused across calls.

Requirements

• Python 3.10+

• A Telegram bot token from @BotFather

Setup

1. Create the bot and get a token

1. Open @BotFather in Telegram and send /newbot.

2. Follow the prompts and copy the token it gives you (looks like 123456789:AA...).

2. Find your chat id

The bot can only message a chat it knows. The simplest path:

1. Send any message to your new bot in Telegram.

2. Ask @userinfobot for your numeric id, or call:

   curl "https://api.telegram.org/bot<YOUR_TOKEN>/getUpdates"

   and read result[].message.chat.id from the JSON.

Set that value as TELEGRAM_DEFAULT_CHAT_ID so the model can send without specifying a chat every time.

3. Install

# with uv (recommended)
uv venv
uv pip install -e .

# or with pip
python -m venv .venv && source .venv/bin/activate
pip install -e .

4. Configure

Copy the example env file and fill it in:

cp .env.example .env
# edit .env: set TELEGRAM_BOT_TOKEN (and ideally TELEGRAM_DEFAULT_CHAT_ID)

Running

The server speaks MCP over stdio, so you normally don't launch it by hand — your MCP client does. To sanity-check that it starts:

TELEGRAM_BOT_TOKEN=... mcp-telegram

It will wait for an MCP client on stdin. Press Ctrl+C to stop.

Connecting a client

Claude Desktop

Add this to your claude_desktop_config.json (~/Library/Application Support/Claude/ on macOS, %APPDATA%\Claude\ on Windows):

{
  "mcpServers": {
    "telegram": {
      "command": "mcp-telegram",
      "env": {
        "TELEGRAM_BOT_TOKEN": "123456789:AA...",
        "TELEGRAM_DEFAULT_CHAT_ID": "123456789"
      }
    }
  }
}

If mcp-telegram is not on your PATH, use the venv's absolute path or run it through uv:

{
  "mcpServers": {
    "telegram": {
      "command": "uv",
      "args": ["run", "--directory", "/absolute/path/to/mcp-telegram", "mcp-telegram"],
      "env": { "TELEGRAM_BOT_TOKEN": "123456789:AA..." }
    }
  }
}

Claude Code

claude mcp add telegram \
  --env TELEGRAM_BOT_TOKEN=123456789:AA... \
  --env TELEGRAM_DEFAULT_CHAT_ID=123456789 \
  -- mcp-telegram

Restart the client after editing its config, then ask the model something like "send me a Telegram message saying the build passed".

Tools

Every send tool accepts an optional chat_id; when omitted it falls back to TELEGRAM_DEFAULT_CHAT_ID. parse_mode is optional and defaults to plain text — only set "MarkdownV2", "HTML" or "Markdown" when you need formatting, and make sure the payload is escaped for that mode.

Development

uv pip install -e ".[dev]"
pytest

Tests mock the Telegram HTTP API with respx, so they run offline and never touch a real bot.

License

MIT — see LICENSE.