Build a Telegram assistant that never sleeps

Telegram is the path of least resistance for putting an AI agent in front of a human. Its bot API is open, a token takes about ninety seconds to get, and there is no review process standing between you and a working bot. That is why almost every agent on this platform ends up with a Telegram front door even when it was built for something else.

The interesting question is not how to connect Telegram. It is what should be behind it.

The part everyone gets right

Getting a bot token is genuinely trivial: message @BotFather on Telegram, send /newbot, pick a name, and it hands you a token. Paste that into the builder when it asks and the bridge does the rest. No webhook to host, no TLS certificate, no public IP.

Worth knowing: the bot is a separate identity from your account. People message it, not you. In a group it only sees messages addressed to it unless you change that, which is usually what you want.

The part that decides whether this is useful

A Telegram assistant that only answers when spoken to is a search box with extra steps. The ones people keep are the ones that act while nobody is watching — and that is a runtime decision, not a channel decision.

So the real question is which of these you are building:

An assistant that handles things overnight

Answers messages, decides what is urgent, and hands you a digest in the morning instead of forty notifications. This needs real scheduling and a memory that spans channels, which means OpenClaw — it runs on the Plus box at $19 a month.

A single-purpose bot

Answers questions about one thing, in one place, well. No scheduler, no sub-agents. PicoClaw does this on the cheapest box we run — $9 a month — and starts faster, which matters because agents here stop when idle.

An engineer you message from your phone

Not a chat bot at all: a coding agent with files, a shell and git that happens to be reachable from Telegram. This is the hosted coding agent shape, from $9 a month.

Building the overnight assistant

Since that is the one most people are actually after, here it is concretely.

1. Write the brief as standing instructions. Not "be a helpful assistant" — say what it should answer on its own, what it should escalate to you immediately, and what it should hold for the digest.
2. Define urgent explicitly. "Anything from my landlord or containing the word 'outage'" is a rule an agent can follow. "Important things" is not.
3. Give the digest a shape. "One message at 8am: what came in, what you answered, what needs me." Without a shape it becomes a wall of text within a week and you stop reading it, which is the same as it not working.
4. Deploy and connect the bot token.
5. Run it for a few days before adding the schedule. Get the answering right first. A cron job on a badly briefed agent is an alarm clock that sends you nonsense.

What it costs to keep up

Hosting is fixed and starts at $9 a month, first agent month free. Model tokens are forwarded at provider cost with no markup, so the variable half depends on volume — and on whether anything is running on a schedule, which is the part that surprises people. See the full cost breakdown.

Two things worth setting up on day one

An explicit "ask me first" list. An agent that answers your Telegram is an agent speaking as you to people who know you. Say plainly what it must never send without checking.

A way to see what it did. Every agent built here reports to ClawMetry, so you get cost per conversation and a full log of what it actually did overnight — and a stop button. For an agent that runs while you sleep, reading the log for the first week is the fastest way to fix the brief.