telegram-slack-mcp

Give Claude your actual chats.

Two MCP servers that connect Claude to your personal Telegram and Slack accounts — your real DMs, groups, and channels, not a bot inbox.

Terminal

claude mcp add telegram -- npx -y telegram-slack-mcp telegram
claude mcp add slack    -- npx -y telegram-slack-mcp slack

Then just talk:

"what's unread on telegram from actual people?" "read my last 30 messages with Priya and draft a reply" "search all my slack workspaces for the deploy postmortem" "what's in the screenshot Priya sent me?"

Why this exists

Most Telegram integrations use the Bot API, which only sees messages sent to a bot you created. It cannot read your existing conversations. telegram-slack-mcp uses MTProto, the same protocol the official app uses, authenticating as you — so Claude sees the chats you actually have.

For Slack, the usual connector is scoped to one workspace. telegram-slack-mcp takes one user token per workspace and treats them as a single surface, so "search everywhere" means everywhere.

Install

Pick whichever runtime you already have. Both ship the same tools.

Node

Terminal

claude mcp add telegram -- npx -y telegram-slack-mcp telegram
claude mcp add slack    -- npx -y telegram-slack-mcp slack

Python

Terminal

claude mcp add telegram -- uvx telegram-slack-mcp telegram
claude mcp add slack    -- uvx telegram-slack-mcp slack

Claude Code plugin

Code

/plugin marketplace add nileshpatil6/telegram-slack-mcp
/plugin install telegram@telegram-slack-mcp
/plugin install slack@telegram-slack-mcp

Other MCP clients (Claude Desktop, Cursor, …)

config.json

{
  "mcpServers": {
    "telegram": {
      "command": "npx",
      "args": ["-y", "telegram-slack-mcp", "telegram"],
      "env": { "TELEGRAM_API_ID": "1234567", "TELEGRAM_API_HASH": "your_hash" }
    },
    "slack": {
      "command": "npx",
      "args": ["-y", "telegram-slack-mcp", "slack"],
      "env": { "SLACK_USER_TOKENS": "xoxp-one,xoxp-two" }
    }
  }
}

Setup

Telegram — 2 minutes

Go to https://my.telegram.org → API development tools → create an app.

Copy the api_id and api_hash, and set them:

Terminal

claude mcp add telegram      -e TELEGRAM_API_ID=1234567      -e TELEGRAM_API_HASH=your_hash      -- npx -y telegram-slack-mcp telegram

Link your account. Either ask Claude — "log into telegram" — or do it yourself in a terminal, which needs no MCP client at all:
bash
```
TELEGRAM_API_ID=1234567 TELEGRAM_API_HASH=your_hash npx -y telegram-slack-mcp login
```
powershell
```
# PowerShell
$env:TELEGRAM_API_ID="1234567"; $env:TELEGRAM_API_HASH="your_hash"
npx -y telegram-slack-mcp login
```
Either way you scan a QR and the session is saved to ~/.chat-mcp.

Through Claude:

It calls login, which returns a link to a page showing a QR code (and the code as text, if your client renders it). Scan it from Telegram on your phone — Settings → Devices → Link Desktop Device — then it calls login_status to confirm. No phone number, no SMS code, no 2FA password typed anywhere. On the same machine you can click "open in Telegram Desktop" instead of scanning.

The session is saved to ~/.chat-mcp and persists; you do this once.

Can't scan? Use the phone-number flow

login_start(phone) sends you a code, login_complete(code, password?) finishes. Note that with this path your phone number and login code pass through the conversation as tool arguments, so they land in the model's context. The QR flow keeps them out of it.

Slack — 5 minutes

Go to https://api.slack.com/apps → Create New App → From a manifest.
Paste slack-app-manifest.yaml — it prefills every scope.
Install to Workspace, then copy the User OAuth Token (xoxp-…, not xoxb-).
Repeat step 3 for each workspace you want. Same app, one token each.
Ask Claude: "connect my slack" — the login tool prompts for the tokens, verifies each one against Slack, and saves them to ~/.chat-mcp. Nothing to put in your config.

Prefer environment variables? That still works:
Terminal
```
claude mcp add slack -e SLACK_USER_TOKENS=xoxp-aaa,xoxp-bbb -- npx -y telegram-slack-mcp slack
```

Slack has no local OAuth flow, which is why you create the app yourself: Slack's OAuth requires a client secret and an HTTPS redirect URL, so a package distributed over npm cannot complete it without either shipping a secret publicly or running a callback server. This project has no server, so it asks for the token instead.

Tools

tool	what it does
`login()`	scan a QR to link your account — no phone number or code
`login_status()`	check whether the QR has been scanned yet
`login_start(phone)`	fallback: send yourself a login code
`login_complete(code, password?)`	finish login, save the session
`whoami()`	which account is connected
`list_chats(limit, query, unread_only, kind)`	recent chats; `kind` filters to `dm`, `group`, `channel`, `bot`…
`read_chat(chat, limit, before_id)`	messages of one chat — `chat` takes @username, id, or part of a name
`search_messages(query, chat?, limit)`	full-text search, one chat or all
`unread_summary(limit, per_chat, kind)`	catch-up view; `kind: "dm"` skips the channel noise
`read_media(chat, message_id, max_kb)`	view a photo or file — returns the actual image, not just its type
`send_message(chat, text, reply_to?)`	reply as you — off unless `TELEGRAM_ALLOW_SEND=1`

Slack

tool	what it does
`login()`	prompts for your `xoxp-` tokens, verifies them, saves them
`whoami()`	every connected workspace and who you are in each
`list_channels(limit, query, workspace, types)`	channels across workspaces
`list_dms(limit, workspace)`	1:1 and group DMs
`read_channel(channel, limit, workspace, before_ts)`	one channel or DM
`read_thread(channel, thread_ts, limit, workspace)`	replies inside a thread
`search_messages(query, limit, workspace)`	supports `in:#chan`, `from:@user`
`unread_summary(limit, per_chat, workspace)`	what's unread everywhere
`read_file(file_id, workspace, max_kb)`	view a file shared in Slack — returns the actual image
`send_message(channel, text, workspace, thread_ts?)`	post as you — off unless `SLACK_ALLOW_SEND=1`

Every Slack tool takes an optional workspace (any substring of the name). Leave it blank to act across all of them; it's only required when a channel name exists in more than one, and the error tells you which.

Configuration

variable	default	meaning
`TELEGRAM_API_ID`	—	from my.telegram.org
`TELEGRAM_API_HASH`	—	from my.telegram.org
`TELEGRAM_ALLOW_SEND`	`0`	`1` lets Claude send Telegram messages as you
`SLACK_USER_TOKENS`	—	one `xoxp-` token per workspace, comma separated
`SLACK_ALLOW_SEND`	`0`	`1` lets Claude post to Slack as you
`CHAT_MCP_DATA_DIR`	`~/.chat-mcp`	where the Telegram session is stored

Sending is off by default in both servers. Reading is the safe default; you opt into writing deliberately.

Privacy

Nothing is sent to any server belonging to this project. There is no such server.

telegram-slack-mcp runs entirely on your machine. The only network traffic it makes is directly to Telegram's and Slack's own APIs, using your own credentials. There is no telemetry, no analytics, no crash reporting, no license check, no "phone home" — grep the source, there is no endpoint to find.

The data path is exactly this:

Code

your machine  <-->  Telegram / Slack APIs      (your credentials, your account)
your machine  <-->  Claude                     (only what you ask Claude to read)

Your credentials never reach Claude. Your session file and xoxp- tokens stay in ~/.chat-mcp and in your environment; the servers use them locally to make API calls and pass back only the messages themselves.

What does reach Claude is the chat content you ask about — because that is the point of the tool. If you ask Claude to read a conversation, that conversation goes into your Claude conversation and is handled under Anthropic's privacy policy, exactly like text you paste in yourself. Nothing else is read, and nothing is read in the background: the servers only act when Claude calls a tool.

The author of this project cannot see any of it.

Security

Read this part.

These act as you, not as a bot. Anything sent shows as sent by your account, and people in those chats cannot tell the difference.
The stored session is a full credential. ~/.chat-mcp/telegram.session and your xoxp- tokens are equivalent to being logged in as you. Anyone who copies them has your account. They never leave your machine, and nothing here phones home.
Everything Claude reads enters the model's context. Don't point this at accounts holding data you would not send to a model.
Telegram calls this a userbot. Automating a personal account is restricted by Telegram's ToS. Light personal read-and-reply use is common and low risk; bulk automation gets accounts banned. Your call, your account.
Revoke any time: Telegram → Settings → Devices, or delete the token in Slack's app settings. Deleting ~/.chat-mcp drops the local session.

Troubleshooting

"Telegram not logged in yet" — run the login flow: ask Claude to log in with your phone number. It persists after that; if it keeps reappearing, check that CHAT_MCP_DATA_DIR is writable.

"Missing Telegram credentials" — TELEGRAM_API_ID / TELEGRAM_API_HASH aren't reaching the server. With claude mcp add, pass them with -e.

Slack returns nothing at all — you're almost certainly authorized against an empty or wrong workspace. Run whoami to see which workspaces are actually connected.

"#general" exists in several workspaces — pass workspace to disambiguate; the error lists your options.

Slack search returns not_allowed_token_type — the token is a bot token (xoxb-). You need the User OAuth token (xoxp-).

Development

bash

git clone https://github.com/nileshpatil6/telegram-slack-mcp
cd telegram-slack-mcp
npm install && npm run build
node dist/cli.js telegram        # run the server on stdio

cd python && python -m build     # build the Python distribution

License

MIT