Skip to main content

​
Set up the Telegram bridge

The Telegram bridge forwards messages between a Telegram bot and the OpenClaw agent running inside the NemoClaw sandbox. You send a message to your bot; the bridge runs it through the agent and returns the response. When the agent needs external network access, the OpenShell TUI prompts for approval. The bridge is an auxiliary service managed by nemoclaw start and nemoclaw stop.

​
Prerequisites

  • A running NemoClaw sandbox, either local or remote.
  • NVIDIA_API_KEY set in your environment.
  • A Telegram bot token from @BotFather.

​
Create a Telegram bot

1

Open BotFather

In Telegram, start a conversation with @BotFather.
2

Create a new bot

Send the /newbot command. BotFather will ask for a display name and a username, then issue a bot token that looks like:
1234567890:AAEhBOweik8ad4b7FelRnTmFqtXRNPEOPio
3

Copy the bot token

Save the token. You will need it in the next step.

​
Configure and start the bridge

1

Set the environment variables

Export the bot token and your NVIDIA API key:
export TELEGRAM_BOT_TOKEN=<your-bot-token>
export NVIDIA_API_KEY=nvapi-...
The bridge script requires both variables at startup. It exits immediately if either is missing.
2

Start auxiliary services

Start the Telegram bridge and the cloudflared tunnel:
nemoclaw start
nemoclaw start launches the following services:
  • Telegram bridge — forwards Telegram messages to the agent and returns responses. Only started when TELEGRAM_BOT_TOKEN is set.
  • cloudflared tunnel — provides external access to the sandbox over a public URL.
3

Verify the bridge is running

Check the status of all auxiliary services:
nemoclaw status
The output shows whether the Telegram bridge and tunnel are running. When the bridge starts successfully, it prints a startup banner:
  ┌─────────────────────────────────────────────────────┐
  │  NemoClaw Telegram Bridge                          │
  │                                                     │
  │  Bot:      @YourBotUsername                        │
  │  Sandbox:  nemoclaw                                │
  │  Model:    nvidia/nemotron-3-super-120b-a12b       │
  │                                                     │
  │  Messages are forwarded to the OpenClaw agent      │
  │  inside the sandbox. Run 'openshell term' in       │
  │  another terminal to monitor + approve egress.     │
  └─────────────────────────────────────────────────────┘

​
Send a message

Open Telegram, find your bot, and send a message. The bridge forwards the text to the OpenClaw agent inside the sandbox and sends the agent’s response back to you. The following built-in commands are handled by the bridge before reaching the agent:
CommandEffect
/startSends a welcome message describing the bot and the sandbox.
/resetClears the active session history for your chat.

​
Restrict access to specific chat IDs

By default, the bridge accepts messages from any Telegram chat. To restrict access to specific users or groups, set the ALLOWED_CHAT_IDS environment variable to a comma-separated list of Telegram chat IDs: 
export ALLOWED_CHAT_IDS="123456789,987654321"
nemoclaw start
Messages from chats not in the list are silently ignored.
To find a chat ID, send a message to your bot and inspect the chat.id field in the Telegram getUpdates response, or use a bot like @userinfobot.

​
Monitor bridge activity

While the bridge is running, open the OpenShell TUI in a separate terminal to monitor sandbox network activity and approve egress requests triggered by the agent: 
openshell term
For a remote sandbox: 
ssh my-gpu-box 'cd /home/ubuntu/nemoclaw && . .env && openshell term'

​
Stop the services

Stop the Telegram bridge and all other auxiliary services: 
nemoclaw stop

​
How the bridge works

When a message arrives, the bridge:
  1. Checks the chat ID against ALLOWED_CHAT_IDS (if set).
  2. Sends a typing indicator to Telegram.
  3. Fetches the sandbox SSH config with openshell sandbox ssh-config nemoclaw.
  4. SSHs into the sandbox and runs openclaw agent --agent main --local -m '<message>' --session-id 'tg-<chatId>'.
  5. Strips setup log lines from the output and returns the agent response to Telegram.
  6. Repeats the typing indicator every 4 seconds while the agent is running.
Long responses are split into 4,000-character chunks to stay within Telegram’s message size limit. Markdown formatting is applied where possible, with a plain-text fallback.
Each Telegram chat uses its own --session-id, so conversation history is isolated per chat. Use /reset to clear a session.

Deploy to a remote GPU

Run NemoClaw on a Brev GPU instance with the Telegram bridge enabled.

Approve network requests

Monitor and approve egress requests from the agent while it responds to Telegram messages.

Build docs developers (and LLMs) love

Get started for freeTalk to us