Telegram Dev

Current official Telegram documentation for Claude Code, Codex and Cursor.

Telegram Dev is an MCP server plus an agent skill. The server fetches, indexes and searches the official docs (Bot API, Mini Apps, MTProto, TDLib, changelogs) and keeps them fresh. The skill teaches the agent to pick the right API, check eligibility and library support, and cite what it read.

Русская версия

Why

Telegram ships a new Bot API version every month or two. A model's knowledge of it stops at its training cutoff, so agents confidently write code against fields that changed or never existed. They also mix up the Bot API, MTProto and TDLib, which have very different capabilities.

With Telegram Dev the agent reads the current page before it writes code. A real Codex CLI run in a fresh project after uvx telegram-dev install --agent codex (tool calls trimmed):

> Use the telegram-dev skill. In a python-telegram-bot project, I want the inline
  'Delete' button to be red and show a custom emoji. Which Bot API fields do this,
  what are the requirements, and what does the request JSON look like?

  telegram_corpus_status {}
  search_telegram_docs {"query": "style icon_custom_emoji_id", "area": "releases"}
  search_telegram_docs {"query": "InlineKeyboardButton icon_custom_emoji_id style", "area": "bot-api"}
  get_telegram_document {"url": "https://core.telegram.org/bots/api", "offset": 167594, "length": 10500}
  get_telegram_document {"url": "https://core.telegram.org/bots/api-changelog", "offset": 39089, "length": 6000}

Use these InlineKeyboardButton fields:
- style: "danger" makes the button red.
- icon_custom_emoji_id is a custom emoji ID shown before "Delete".
- callback_data is your action identifier, 1-64 bytes.

The custom emoji requires either a bot that bought additional usernames on Fragment,
or a Premium bot owner with the message sent by the bot to a private chat, group or
supergroup. PTB 22.7+ supports both fields. Older Telegram clients show the button
unstyled. Source: https://core.telegram.org/bots/api#inlinekeyboardbutton

{"chat_id": 123456789, "text": "Choose an action:",
 "reply_markup": {"inline_keyboard": [[{"text": "Delete", "callback_data": "delete:42",
   "style": "danger", "icon_custom_emoji_id": "YOUR_CUSTOM_EMOJI_ID"}]]}}

Install

You need uv. No bot token, Telegram account or API key is needed: the server only reads public documentation.

MCP server and skill, one command

Run this in your project directory:

uvx telegram-dev install --agent claude    # or: codex, cursor

The installer keeps your other settings, backs up the files it changes and never overwrites an existing Telegram Dev setup. The server is pinned to the installed version, so nothing has to stay on disk besides the uv cache. Open a new agent session afterwards and approve the project MCP server when asked. Codex loads project settings only after you trust the project directory.

To upgrade later: uvx telegram-dev@latest install --agent claude --update.

Claude Code plugin

/plugin marketplace add Kay0k1/telegram-dev
/plugin install telegram-dev@telegram-dev-tools

This adds the skill and the MCP server for all your projects. Use either the plugin or the installer, not both.

MCP server only

claude mcp add --scope user telegram-knowledge -- uvx telegram-dev serve

codex mcp add telegram-knowledge -- uvx telegram-dev serve

Cursor:

Any other MCP client:

{
  "mcpServers": {
    "telegram-knowledge": { "command": "uvx", "args": ["telegram-dev", "serve"] }
  }
}

The tools work without the skill, but the skill is what makes an agent use them consistently.

Tools

The cache starts empty. Pages are fetched the first time the agent reads them and indexed right away. To pre-fetch everything reachable from the configured sources:

uvx telegram-dev sync --max-pages 0
uvx telegram-dev search "sendVideoNote"
uvx telegram-dev status

What the skill covers

The skill lives in skills/telegram-dev, with one reference file per area.

How fresh the docs are

• Reading a page older than 6 hours re-fetches it. Search re-fetches up to six stale matching pages.
• While the server runs, a background cycle checks up to 120 pages every 6 hours: entry points, newly found links, oldest pages.
• If Telegram's site is unreachable, the cached text is kept and marked stale with the error, so the agent knows it is reading an old copy.
• A daily GitHub Action checks all 52 official sources and publishes hashes and errors on the docs-status branch.

Settings such as the cache path (TELEGRAM_KNOWLEDGE_DB) and refresh intervals are in docs/runtime.md.

Self-hosting

The server also runs over Streamable HTTP, so a team can share one cache. A Dockerfile and a Caddy setup with HTTPS are included: docs/deployment.md.

uvx telegram-dev serve --transport streamable-http --port 8000

Limits

• Search is local full-text search (SQLite FTS5), not semantic search. Exact method names and English terms work best. A small set of Russian aliases is built in.
• Coverage is what can be reached from the 52 official entry points. Page history starts when your installation first fetched the page.
• The server does not log in to Telegram, send messages or touch chats. It only reads public pages.
• The skill improves how agents work. It does not guarantee a correct answer: check the code against Telegram before shipping.

Development

git clone https://github.com/Kay0k1/telegram-dev && cd telegram-dev
uv sync --frozen
uv run --frozen python -m unittest discover -s tests -v
uv run --frozen telegram-dev install --agent claude --project /path/to/test-project

CI runs the tests on Linux, macOS and Windows and checks the Docker image. docs/validation.md lists what has been verified by hand and what has not. Issues and pull requests are welcome; tests/behavioral-cases.json has agent scenarios that still need runs in each client.

License

MIT for the code and the skill. Fetched Telegram documentation remains Telegram's; the package does not bundle it (NOTICE). This is an independent project, not affiliated with Telegram.