Telegram AI bot with Claude — a 5-minute self-hosted setup

Telegram is the cleanest bot platform on the consumer internet. You don’t need to verify a business, pay for a Cloud API, or expose a webhook URL. Five minutes from “I want a Telegram AI bot” to “it works”.

This is how to do it with openclawOS — self-hosted, your model of choice, full Pi agent capabilities.

Step 1: Install openclawOS

npm install -g openclawos@latest
openclawos onboard --install-daemon

The Control UI opens on localhost:7777.

Step 2: Create a bot

Message @BotFather on Telegram. Send /newbot. Pick a display name and a username ending in bot. BotFather hands you a token that looks like 123456789:ABCdefGhIJKlmnopqRstuv.

Copy it.

Step 3: Pair Telegram

openclaw apps install telegram

The CLI prompts for the token. Paste, hit enter. Or do it from the Control UI’s Channels tab.

Step 4: Drop in your LLM key

In the Control UI’s Providers tab, paste your Anthropic API key. We recommend Claude Opus 4.6 for Pi’s full tool-use behaviour; Sonnet 4.6 also works and is cheaper.

If you’d rather use OpenAI, Gemini, or a local Ollama, those work too — pick in the dropdown.

Step 5: Say hi

Open your bot in Telegram. Send /start. Pi responds. Try:

• “What’s on the front page of Hacker News?”
• “Summarise this PDF” (attach a file)
• “Remember that I’m vegetarian and based in Berlin”

You’re done. Total time: under 5 minutes assuming you didn’t refill your coffee.

What you can do from here

Commands Pi auto-registers

• /new — start a fresh session (forget current memory)
• /fork — branch the current session
• /summary — compact the current session
• /agents — list available agents (if more than one)
• /help — Pi explains what it can do

These all show up in BotFather’s command list if you choose to publish them.

Bind to a group

By default Pi only listens to DMs. To make it active in a group:

# ~/.openclaw/bindings/telegram-team.yaml
trigger:
  channel: telegram
  scope: { groupId: "-1001234..." }
  match:
    mentions: ["@your_bot"]
agent: pi
session:
  scope: group

Now @your_bot what did we discuss yesterday? triggers Pi in that group.

Inline mode

In your bot’s settings on @BotFather, enable inline mode. Now in any chat, type @your_bot something — Pi will offer inline results you can tap to send.

Inline is great for summaries, code snippets, structured replies you want to drop into another conversation.

Multiple bots

Pair as many bots as you like. Each one is its own binding. Common setups:

• @dipankar_dev_bot → coding Pi, listens in your engineering Discord too
• @dipankar_personal_bot → personal Pi, also handles your iMessage
• @dipankar_research_bot → a research Pi with a different system prompt and different tools enabled

Cost expectations

A Telegram-Pi chatting with you a few dozen times a day costs $3-10/month in Claude Sonnet tokens. Opus is 5-10× more. Local models are free per token but the round-trip is slower and tool-use reliability is lower.

Privacy and rate limits

Telegram bot DMs are server-side; Telegram does see them as a vendor (they always have for any bot). The openclawOS Gateway sits between Telegram and your LLM — Pi processes only the latest turns, not your full history.

Rate limits: Telegram caps bots at 30 messages/second per chat, 20 messages/minute per group. Pi’s tool use never approaches this in normal use.

Beyond the basic loop

The fun starts when you wire bindings together. A common second move is connecting Pi to a vector store of your own notes, so you can ask Telegram-Pi about your Obsidian vault. Or wire a cron trigger to send you a morning brief on Telegram every day at 7am. Or expose Pi as a webhook so your CI can ping it when a build fails.

The Gateway handles all of this. Telegram is just the surface.