Do I need a server to run a Telegram bot?

Yes — the bot is a long-running process that either polls Telegram for updates or receives webhooks. A small always-on instance on a free tier (Railway, Fly.io, Render) is enough. Your OpenAI key lives there as an environment variable, never in the client.

What is the difference between polling and webhooks?

Polling has your bot repeatedly ask Telegram "any new messages?" and is the simplest to start with. Webhooks have Telegram push updates to a public HTTPS URL you expose, which scales better and uses fewer resources. Begin with polling, switch to webhooks for production.

How does the bot remember a conversation?

Telegram delivers each message independently, so you maintain a dictionary keyed by chat ID holding that chat's message history. You resend the relevant history to the model on each turn, then append the reply. Trim old turns as you approach the context limit.

Can it handle voice messages?

Yes. Telegram sends voice notes as OGG audio files you download via the bot API, transcribe with a speech-to-text model like Whisper, then treat the transcript as a normal text turn. The reply can be sent back as text or synthesized speech.

How much does it cost to run?

The Telegram Bot API is free. Your only real cost is the LLM usage, billed per token of input and output. A personal assistant bot for one user typically costs a few cents to a couple of dollars a month — estimate it with a token cost calculator before scaling to many users.

How to Build an AI Telegram Bot

A Telegram bot is one of the most satisfying first AI projects: it lives in an app you already check constantly, it works on every device, and Telegram handles all the messaging plumbing for free. This tutorial walks through creating the bot, connecting it to an LLM with conversation memory, and adding voice support — plus an interactive generator that produces ready-to-run Python starter code.

Step 1 — Create the bot with BotFather

Every Telegram bot is registered through @BotFather, an official bot that issues credentials. Open a chat with it, send /newbot, choose a name and a unique username ending in bot, and BotFather replies with an HTTP API token that looks like 123456:ABC-DEF.... That token is your bot’s password — keep it in an environment variable, never in source control.

Step 2 — Connect to the LLM with memory

Install python-telegram-bot and register a message handler — a function called for every incoming text message. Inside it you forward the user’s message to the OpenAI Chat Completions endpoint and reply with the result.

The key detail is memory. Telegram delivers each message as an isolated update, so you keep a dictionary keyed by chat_id, each holding that chat’s running list of role-tagged messages (system, user, assistant). You resend that history on every call — it is the bot’s memory — and append the reply before the next turn. When the history grows near the model’s context limit, drop or summarize the oldest turns.

Step 3 — Add voice, then deploy

To accept voice notes, add a handler for the VOICE filter: download the OGG file Telegram provides, transcribe it with a speech-to-text model such as Whisper, and feed the transcript in as a normal user turn. The reply can go back as text or as synthesized audio.

For deployment, start with polling (the bot asks Telegram for updates in a loop) because it needs no public URL. Push the process to a free always-on host like Railway, Fly.io, or Render, set your tokens as environment variables, and you have a personal assistant in your pocket. Switch to webhooks later for lower latency and cost at scale.

Use the generator below to produce a starter bot.py, then read how to build a chatbot with the OpenAI API for deeper context handling and estimate spend with the LLM API Cost Calculator.