Telegram AI Assistant: Run Your Own Agents From Your Phone

Two different products wearing one name

The first class is a chatbot that happens to live in Telegram. The open-source bridge projects are the honest, DIY version of it: father-bot/chatgpt_telegram_bot (Python, MIT license, 5,527 GitHub stars, still active as of June 2026) and n3d1117/chatgpt-telegram-bot (Python, GPL-2.0, 3,463 stars, quiet since mid 2025) both wire an LLM into a Telegram chat using your own API key (an API is the interface programs use to talk to a service; the key is what bills it to you). Hosted products like TeleClaw go further and build the bot for you: a free tier, then paid credits from $5 up to $2,000 a month, with the agents running on TeleClaw's infrastructure. Credit where due, all of these do what they say. They put a capable AI inside a chat window, cheaply.

But read what their own pages describe. As of 2026-09-02, the two bridge projects' READMEs document chat: answers, images, voice. They do not document controlling agents that run on your own machines. TeleClaw's homepage does not document self-hosting or running on your own hardware at all; it is hosted cloud as documented. And neither AI vendor ships a consumer chatbot here: as of the same check, OpenAI lists no official ChatGPT Telegram bot (the popular options are community projects using the API), and Anthropic's documented Telegram offering is not a chatbot either. It is something more useful.

The second class: a Telegram bot as the remote control

In the second class, the intelligence does not live in Telegram at all. It lives in an agent system on hardware you control, and Telegram is only the wire to it. Three current, documented examples, from different vendors, which is what tells you this is a pattern and not one product's trick:

Anthropic's official Claude Code Telegram plugin is an MCP server (Model Context Protocol, a standard plug for wiring tools into an AI session) that logs into Telegram as a bot (an automated account that a program controls) and forwards your direct messages into a running Claude Code session on your machine. The session can reply, attach files up to 50MB, react with Telegram's fixed emoji set, and edit its own messages. Setup is one BotFather token (the credential Telegram's bot-registration account hands you, explained below) and a configure command. It is a bridge to your own session, not a hosted assistant.

OpenClaw is an open-source personal assistant that runs locally on macOS, Windows, or Linux: inbox and calendar, web browsing, file read and write, shell commands, persistent memory. It is free, MIT licensed, governed by the non-profit OpenClaw Foundation, and it has the numbers to prove people use it: 388,631 GitHub stars as of 2026-09-02 on a repo created in November 2025. Telegram is one of 29 chat channels it documents. Its Telegram docs describe exactly the shape this article is about: a BotFather token, long polling as the default transport (your machine asks Telegram for new messages, so no public server is needed; both terms are unpacked below), and access control before anything else.

n8n, the workflow automation tool, ships a built-in Telegram node and trigger, so a self-hosted workflow can receive your message and act on it. Its template gallery includes DIY personal assistants built this way, like one named Angie with Telegram voice and text.

Note what these have in common and the first class lacks: the thing doing the work is yours, it persists, and Telegram never sees more than the conversation. This is the same buy-or-build fork I mapped in the assistant-with-memory guide, showing up one layer down.

How the pattern works, from zero

The pleasant surprise is that the Telegram half costs nothing and needs no server. Bots are free to create and free to run; Telegram's platform hosts more than 10 million of them. You message @BotFather inside Telegram, send the /newbot command, and get back a token (a password-like string; whoever holds it has full control of your bot, so treat it like a house key and store it accordingly).

Then your software has two mutually exclusive ways to receive messages. A webhook means Telegram pushes each message to a public HTTPS address you host, which most home setups do not have. Long polling means your machine repeatedly asks Telegram whether anything new arrived, which works from behind a home router with no public address, no port forwarding, nothing exposed. This is why the pattern is honest for a beginner: long polling is the default transport in OpenClaw's Telegram channel, and it means the machine under your desk can serve a bot the same afternoon you create one.

Locking it down, from the docs, not from fear

The scope facts come straight from Telegram's own documentation, and they mostly cut in your favor. A bot cannot start a conversation; a human has to message it first or add it to a group. In groups, privacy mode is on by default, so the bot sees commands addressed to it rather than everything said. Sending is rate limited on the free tier: roughly one message per second per chat, 20 per minute into a group, about 30 per second for broadcast-style sends. For a personal assistant sending a handful of messages a day, these limits are not the constraint.

The part you must handle yourself is who may talk to your bot, because the bot's Telegram username is findable. OpenClaw's docs treat this as the first problem: the default DM policy is pairing, where a new user gets an approval code that expires after an hour, and the stricter option is an allowlist (a list of the only numeric user IDs the bot will answer). Groups get a separate policy layer. Whatever harness you use, the rule is the same: the one person Telegram will relay messages from should be you, and nobody else.

One more documented fact matters more than it looks: the Bot API exposes neither message history nor search to the bot. Anthropic's plugin README states this plainly as the reason no history tool exists. The bot sees messages only as they arrive, and Telegram stores undelivered updates for at most 24 hours. The durable record of everything you and your assistant said lives in your own Telegram account, on your phone, outside your system entirely.

What that felt like the day it mattered

That last fact is not a footnote for me. When a configuration reset wiped this factory's workspace in February 2026, the most complete record of the system was the Telegram chat history: every message I had sent while building, with screenshots, sitting in an app the reset could not reach. The remote control turned out to be the black box recorder.

Day to day the pattern is quieter than that: message the assistant from wherever you are, and the machine at home does the work and answers back. No extra app, no dashboard, no laptop in the bag. The messenger already on your phone becomes the front door to a system you own, and building that system is the actual work. The bot token is the easy part.