How to run Claude as your personal Telegram assistant: the actual stack
You can't download Claude's weights, so "self-hosting Claude" in the literal sense isn't a thing. But the thing most people actually want — Claude as a personal Telegram assistant that lives on their own server, remembers them, and works around the clock — is entirely buildable. Here's the real stack, piece by piece, and where each part runs.
The short answer first: to run Claude as a personal Telegram assistant, you use Claude as the brain through Anthropic's API (or a subscription-backed CLI), create a Telegram bot with BotFather to get a token, and run a small program on a server you own that connects the two — it receives your Telegram messages, calls Claude, and sends the answer back. On top of that wiring you add a memory database so it remembers you, and skills and integrations so it can actually do things. That's the whole shape of it. Everything below is just filling in each piece so you know what you're building before you build it.
What "running Claude as your Telegram assistant" actually means
It helps to separate two ideas people blur together. Self-hosting the model would mean downloading Claude's weights and running inference yourself — and that's impossible, because Anthropic doesn't release the weights. We covered exactly why in can you self-host Claude, and the honest answer. Running Claude as your assistant means something different and achievable: Claude stays where it is, behind Anthropic's API, and you own everything around it — the Telegram bot, the server, the memory, the skills. The reasoning is rented; the assistant is yours. That distinction is the entire point, and getting it wrong is why people either give up ("you can't self-host Claude") or overpay someone selling a fantasy.
How to run Claude as your personal Telegram assistant, step by step
The build is five parts, and each one adds a capability the previous one lacked. Here's the stack laid out, with what each step gives you and where it physically runs:
| Step | What you do | Where it runs |
|---|---|---|
| 1. The brain | Get an Anthropic API key, or use a Claude subscription through the Agent SDK / CLI | Claude API (stateless) |
| 2. The face | Create a bot with @BotFather in Telegram, copy the bot token | Telegram |
| 3. The wiring | Run a small program (the Claude Agent SDK) that passes messages between Telegram and Claude | Your server |
| 4. The memory | Add a database that stores facts, preferences and past conversations, recalled each message | Your server (database) |
| 5. The hands | Add skills / integrations — calendar, email, reminders, browsing, posting | Your server + external APIs |
Steps 1–3 give you a working Claude chatbot inside Telegram. Steps 4 and 5 are what turn that chatbot into an actual assistant — one that remembers you and can act. Skip them and you've built a slightly nicer chat window; include them and you've built something that works for you while you're not looking. The skills and integrations layer is where most of the "assistant" feeling actually comes from.
Subscription vs API key — how you pay for the brain
Both power the same assistant, but they bill in opposite ways, and choosing wrong is the most common way to get an unpleasant surprise. An API key charges per token: clean and predictable for light use, but a personal assistant you talk to all day — plus one that thinks in the background — can run up real cost. A subscription-backed setup, driven through the Claude Agent SDK, uses your plan's included usage instead of metering every call, which is usually the cheaper path for an always-on assistant. One hard rule: don't have both configured at once — an API key present in the environment overrides the subscription and silently puts you back on per-token billing. Decide which one is powering the brain, and make sure only that one is set.
Where it runs — and why that's the whole privacy story
An assistant on your laptop dies the moment the lid closes, so for anything "personal assistant" implies — reminders that arrive, jobs that run overnight — it needs to live on a server that's always on. That's usually a small VPS for a few dollars a month; the walkthrough is in how to deploy a personal AI assistant on a VPS. Putting it there is also what makes it a 24/7 assistant rather than a chat you have to open. And it's the privacy story too: the program, the token and the memory database all sit on hardware you control, and only the individual prompt travels to Claude's API for each reply — nothing about your history is retained on the provider's side. That's the same architecture-not-policy point behind a self-hosted ChatGPT alternative: privacy comes from where your data lives, not from a checkbox.
Where it stops being a chatbot
The gap between "Claude in Telegram" and "my assistant" is memory plus skills. Memory is what lets it recognise you on the first message of a fresh conversation instead of re-introducing yourself daily. Skills are the integrations that let it act — check your calendar, send an email, set a reminder, run a task on a schedule. Claude supplies the reasoning, the memory supplies continuity, and the skills give it hands. This is also the honest reason to build it as a real assistant rather than clone a generic bot: a model-agnostic "put any AI in Telegram" tutorial (we wrote one — how to build a personal AI assistant on Telegram) stops at the wiring. Running Claude as your assistant, with memory and skills on your own server, is the version that keeps being useful past week one.
The cost of getting this wrong
Two failure modes, both quiet. The first is building only steps 1–3 — a Claude chatbot in Telegram with no memory and no skills — and then wondering why it never feels like an assistant: it can't remember you and it can't do anything, so it's just ChatGPT with a different icon. The second is the billing trap: leaving an API key set "just in case" alongside a subscription, and discovering weeks later that every message was metered per token the whole time. Avoid both by treating this as a five-part stack, not a one-line hack — decide your billing path up front, and don't stop at the wiring. Do that and you get the thing you actually wanted: Claude, as your assistant, on your terms.
The verdict
You can't self-host the model, but you can absolutely run Claude as a personal Telegram assistant that's yours end to end — Claude for the reasoning, a BotFather bot for the face, a small program on your own server for the wiring, a memory database for continuity, and skills for the hands. Get the billing path right and it stays cheap; keep it on your own server and it stays private and always-on; add memory and skills and it stops being a chatbot. That exact stack — Claude as the brain, everything else owned by you — is what Avelina AI ships as a single install. See what the assistant can actually do on the skills and integrations page.
FAQ
Can you run Claude as your own Telegram assistant?
Yes — Claude stays behind Anthropic's API as the brain, and you connect it to a BotFather bot via a small program on your own server. You own everything except the model itself. More on the limits in can you self-host Claude.
API key or subscription?
Either works; a subscription-backed setup is usually cheaper for an always-on assistant, an API key is predictable for light use. Never set both — the API key overrides the subscription.
Where does it run?
On a server you control — a small VPS keeps it online 24/7. Only the prompt goes to Claude's API; your memory stays on your side.
What makes it a real assistant?
Memory and skills on top of the wiring — continuity plus the ability to actually do things.