tg-mcp
MCP server that gives MCP clients access to a personal Telegram account. It can read chats, search messages, download media, and send messages or files while keeping credentials local.
README
tg-mcp
MCP server for a personal Telegram account. Read, search, download media, and send messages or files, from any MCP client.
Credentials stay on the machine in ~/.tg/ and are never returned by any tool.
Setup on a new machine
Needs uv and nothing else. Python, dependencies and the package are all fetched on demand.
1. Sign in (once per machine, interactive because Telegram sends a code):
uvx --from git+https://github.com/Abdk4Moura/tg-mcp tg-mcp-login
It prompts for api_id / api_hash from https://my.telegram.org (API development
tools), then your phone number, then the code Telegram sends to the app. It writes
~/.tg/credentials.json and ~/.tg/session.session, both mode 600, and prints your
recent chats with their ids.
2. Register the server:
claude mcp add telegram -- uvx --from git+https://github.com/Abdk4Moura/tg-mcp tg-mcp
That is the whole setup. claude mcp list should show telegram ✔ Connected.
Setting it up for other agents
The sign-in is per machine, not per agent. Run tg-mcp-login once and every agent on
that machine shares ~/.tg/. Adding a second or third client is config only.
There is no single config format across MCP clients, so this package ships an installer that writes each one in its own shape:
uvx --from git+https://github.com/Abdk4Moura/tg-mcp tg-mcp-install
It configures every supported client it finds and leaves the rest alone. --list shows
what it detected, --dry-run prints the changes without making them, --name sets the
server name, and naming clients (tg-mcp-install opencode cursor) limits it to those.
Existing config is merged rather than replaced, a .bak is kept, and a config file it
cannot parse is skipped with a warning rather than overwritten.
| Client | How it is registered |
|---|---|
| Claude Code | claude mcp add |
| Codex | codex mcp add |
| opencode | ~/.config/opencode/opencode.json |
| Claude Desktop | claude_desktop_config.json |
| Cursor | ~/.cursor/mcp.json |
| Windsurf | ~/.codeium/windsurf/mcp_config.json |
The rest of this section is the same thing by hand. Under every format it is one command and three arguments:
| command | uvx |
| args | --from, git+https://github.com/Abdk4Moura/tg-mcp, tg-mcp |
Claude Code
claude mcp add telegram -- uvx --from git+https://github.com/Abdk4Moura/tg-mcp tg-mcp
Add --scope user to make it available in every project rather than just the current one.
Claude Desktop, Cursor, Windsurf, Continue, and most others
These share the same mcpServers block. Add:
{
"mcpServers": {
"telegram": {
"command": "uvx",
"args": ["--from", "git+https://github.com/Abdk4Moura/tg-mcp", "tg-mcp"]
}
}
}
to the client's config file:
| Client | File |
|---|---|
| Claude Desktop (macOS) | ~/Library/Application Support/Claude/claude_desktop_config.json |
| Claude Desktop (Windows) | %APPDATA%\Claude\claude_desktop_config.json |
| Cursor | ~/.cursor/mcp.json, or .cursor/mcp.json for one project |
| Windsurf | ~/.codeium/windsurf/mcp_config.json |
| Continue | ~/.continue/config.json |
Restart the client afterwards. Most only read MCP config at startup.
opencode
opencode does not use mcpServers. Its key is mcp, the command is a single array rather
than a command plus args, and each server declares its type. Add to
~/.config/opencode/opencode.json (or opencode.json / .opencode/opencode.json in a
project, for that project only):
{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"telegram": {
"type": "local",
"command": ["uvx", "--from", "git+https://github.com/Abdk4Moura/tg-mcp", "tg-mcp"],
"enabled": true
}
}
}
Environment variables go in an environment object, not env. opencode mcp list shows
the server and whether it connected.
opencode reads opencode.json and opencode.jsonc and merges them, so if your config is
a commented .jsonc you can drop the block above into a plain opencode.json beside it
and leave the commented file alone. That is what tg-mcp-install does.
opencode mcp add telegram also works, but it prompts for the command interactively, so
it is not scriptable for a local server the way claude mcp add is. The --url flag on
that command is for remote servers only.
Codex
codex mcp add telegram -- uvx --from git+https://github.com/Abdk4Moura/tg-mcp tg-mcp
Codex stores this in ~/.codex/config.toml. codex mcp list shows it.
Anything else
Give the client the same command and args from the table at the top of this section, in whatever shape its own MCP documentation specifies.
If uvx is not on PATH
Some clients launch with a minimal environment and will not find uvx. Use its absolute
path as the command (which uvx to find it, commonly ~/.local/bin/uvx).
Tools
| Tool | What it does |
|---|---|
tg_whoami |
Which account is signed in |
tg_list_chats |
Recent chats with their numeric ids |
tg_read_chat |
Read messages; peer takes an id, @username, or me for Saved Messages; optional since / until as YYYY-MM-DD |
tg_search |
Search message text across all chats or within one peer |
tg_download_media |
Fetch a message's attachment to disk |
tg_send_message |
Sends text. Real, immediate, not recallable by this server |
tg_send_file |
Sends a file with an optional caption |
Sessions
Each machine that runs tg-mcp-login creates its own Telegram session, listed
separately under Settings → Devices and revocable on its own.
Copying session.session between machines also works, but both then share one session
identity: revoking it logs out both, and a login from a very different IP can trip
Telegram's security checks. Prefer a fresh login per machine.
That file is full account access. Anyone holding it can read and send as you.
Configuration
| Variable | Default | Purpose |
|---|---|---|
TG_DIR |
~/.tg |
Where credentials.json and session.session live |
TG_DOWNLOAD_DIR |
$TG_DIR/downloads |
Where tg_download_media writes |
Local development
uv run --directory /path/to/tg-mcp tg-mcp
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。