Robin

Problem

Assistant apps want you to live inside them. I already live inside Telegram, and the things I actually lose track of - a task mentioned in passing, a follow-up, a reminder for tomorrow morning - happen in the middle of a conversation, not in a separate app I would have to remember to open.

Approach

Put the assistant where the messages already are. Robin runs as a Telegram bot on OpenClaw, with custom Python tools for the things I need it to do and a Llama model reading intent from plain language. Both English and Bengali. It is hosted on DigitalOcean rather than a managed assistant platform, so the tools, the model and the data stay mine.

The core journey

1. 01Message the bot in plain language
2. 02Llama reads intent from the message
3. 03OpenClaw routes it to a registered tool
4. 04Custom Python tool performs the action
5. 05Reply comes back in the same chat

Custom tools, not a chat wrapper

The value is not the conversation, it is what the conversation can reach. Capabilities are registered as Python tools - capture a task, plan the day, set a reminder, draft a message, organise notes, track a follow-up - and the model picks the tool rather than producing text about it.

Device pairing over a private network

A /pair flow connects a device to the gateway. The gateway binds to loopback by default and is reached over Tailscale rather than being published to the open internet, which means the assistant can touch a machine without that machine having a public address.

Outcome

Running and in daily use. It is the first thing I built that I use every day without meaning to test it - the point at which a project stops being a demo.

What it taught me

• Meeting people in the tool they already use beats building a better tool they have to adopt.
• Binding a gateway to loopback and reaching it over a private network is a better default than exposing a port.