svoy — svoy — home

Hi, this is svoy: a personal AI agent that lives in your Telegram

We're used to AI being a tab. Open it, explain who you are and what you're working on, paste the same context, get an answer, close it. It's useful. So is a hammer.

An agent that lives in your Telegram feels different. You don't "open" it — you text it. The way you text friends, in the same chat list, between your mum and the work group. A voice note on the go, a forwarded contract, "remind me on Friday". It answers, remembers what you talked about last Tuesday, and sometimes writes first: with what matters today, or with what it got done while you slept.

It isn't that it's smarter. Models are roughly at the same level everywhere now. It's that it's there. Not an app you launch, but someone who's already around.

And that it's yours. I want to explain that word without the marketing, because everything rests on it.

Its own computer.
Your agent gets its own isolated Ubuntu machine with a persistent disk, separate from everyone else's. That's where it runs code, keeps files, searches the web and holds on to what you give it.
Its own memory.
Everything it knows about you lives on that machine's disk, as plain files and a database of your conversations. You can open, read, correct, download or delete them. Your memory doesn't dissolve into someone else's cloud or turn into someone's training set.
Your own bot.
The agent lives in a Telegram bot that belongs to you: your name for it, your avatar, your account as the owner. Not one shared bot for a million people where you're a row in a table.

Under the hood is Hermes Agent, an open-source project under the MIT license. I didn't hide it or rename it "our proprietary technology". If you ever want to leave, take its memory and skills and run the same Hermes anywhere: your own server, your laptop. That seems fair to me. You should stay because it's good, not because you're locked in.

Now, what isn't only yours, so there are no asterisks. To think, the agent uses a language model: GLM by Z.ai, through OpenRouter. The text of a request goes there and comes back as an answer. When the agent searches the web, the query goes to a search engine, and when it answers with a voice note, Microsoft's service reads the text aloud. Messages to the bot travel through Telegram's servers, like any other chat. And the machine runs in a data centre that I rent. I don't read people's conversations and don't intend to, but trust is part of this deal, and it's better to say so up front.

So why me, if you can install Hermes yourself? You can. It's an evening of VPS, Docker, BotFather, keys and logs, and after that the updates and the Saturday crashes. I wanted it to take one button for someone who has never opened a terminal. Press it, and a minute later someone of your own says hello.

For now it's a beta: free, with a model budget of $10 a month. When there's a price, I'll write it right here, in one line.

I'm not building another productivity tool or another chat with a neural network. There are plenty of those, and many are good. I'm making sure there's someone of your own in your phone, and that everything it knows about you belongs to you.

— the maker of svoy

Our bot opens. A couple of questions, and your agent is ready.

There will be no screenshots. Here is one of its Tuesdays:

This is one ordinary day. Scroll, or press "Next". Tap any message to see what the agent was doing on its computer at that moment.

What to show

Biscuit

bot

  1. Biscuit: Morning. Your call with Marina is at 11; you asked me to remind you. Rain after lunch, take an umbrella. And the internet bill still isn't paid, it's due tomorrow.
    Under the hood
    • Schedule · weekdays 07:30 — morning brief
    • Memory · "call with Marina, Tue 11:00"
    • Search · weather forecast
  2. You: Voice message, 0:14
    Under the hood
    • Voice → text · 14 s, right on the machine
  3. Biscuit: 🎙️ "Find a cabin for the weekend, under 200 a night, with a sauna, and it has to be OK with dogs"
  4. Biscuit: Got it: a cabin for the weekend, under $200 a night, with a sauna, dog-friendly. I'll look and send options by lunch.
    Under the hood
    • Schedule · today 13:00 — find a cabin
  5. You: File: Lease_agreement.pdf412 KBanything sketchy in here?
    Under the hood
    • File · ~/docs/Lease_agreement.pdf
  6. Biscuit: Three things I'd raise. Clause 4.2: leaving early costs two months' rent. Clause 6.1: the landlord can raise the rent 'at their discretion', with no cap. And there's no inventory at all, so there's nothing to point to if the washing machine breaks. The rest is standard. I saved a copy.
    Under the hood
    • Terminal · pdftotext, 11 pages
  7. Biscuit: Found three. 1) Pine Hollow, $160, wood-fired sauna, dogs welcome. 2) A lake house with its own dock, $190. 3) A small cabin in the woods, $120, but the sauna is shared. I like the second one.
    Under the hood
    • Schedule · ran at 13:00
    • Web search · 9 queries, 41 options → 3
  8. You: take the second. and keep an eye on the price
  9. Biscuit: Watching it. Every night until Friday.
    Under the hood
    • Schedule · nightly until Fri — lake house price
  10. You: every friday at six tell me what I spent on taxis that week. I'll drop the receipts here
  11. Biscuit: Deal: Fridays at 6pm. I'm keeping the receipts in taxi.csv on my disk; ask and I'll send you the whole table.
    Under the hood
    • Schedule · Fri 18:00 — taxi spend
    • File · ~/data/taxi.csv created
  12. You: what did I say about a gift for Lena?
  13. Biscuit: On September 17 you wanted to get her a film camera, 'a proper one, not a disposable'. Her birthday is November 2, three weeks from now. Want me to look at options?
    Under the hood
    • Conversation search · "gift Lena" → September 17
  14. You: write a script that renames photos by the date they were taken
  15. Biscuit: File: rename_by_date.py2 KBTested it on twenty sample photos on my machine, it works. Run: python3 rename_by_date.py path/to/folder. I've saved how to do this, so next time it'll be quicker.
    Under the hood
    • Terminal · 20 files — ok
    • File · ~/scripts/rename_by_date.py
    • Skill created · "Rename photos by date taken"
  16. Wednesday
  17. Biscuit: Morning. The lake house dropped by $20, it's $170 now. Book it for Saturday?
    Under the hood
    • Schedule · nightly price check
    • Web search · 3 queries

Biscuit's computer

Online
Parts of the computer
  • USER.mdwhat it knows about you
  • MEMORY.mdits notes about the work
  • History1,248 messages

59 · latest: "Rename photos by date taken", today 21:32

  • Rename photos by date takenwrote it itself · today 21:32
  • Weekdays, 07:30morning brief
  • Nightly until Frilake house price
  • Fri, 18:00taxi spend for the week
  • ~/docs/Lease_agreement.pdf412 KB
  • ~/data/taxi.csv1 KB
  • ~/scripts/rename_by_date.py2 KB
Doesn't work in the demo — it will for you

Ubuntu 24.04 · 1 vCPU · 1.5 GB · 10 GB disk

That was an ordinary Tuesday. By Wednesday it knows you a little better.

That's the whole product, really: a bot where someone of your own answers you, and the computer it works on. It will show you the rest itself — getting acquainted takes a minute.

Our bot opens. A couple of questions, and your agent is ready.

What follows is the long, calm version: for search engines, and for people who like to know exactly how things work.

A personal AI agent in your Telegram

svoy is a managed service that gives you a personal AI agent in Telegram with one button. Every agent gets its own isolated Ubuntu machine with a persistent disk, its own memory and its own Telegram bot. Under the hood it runs Hermes Agent, the open-source agent by Nous Research, released under the MIT license. The model is GLM by Z.ai, connected through OpenRouter. You don't set up a server, API keys or a terminal: open the bot, press the button, give your agent a name, and a minute later it's there.

Agent vs chatbot

A chatbot answers a message. An agent works on a task. It has tools (a terminal, files, web search), decides which ones it needs, takes several steps in a row and comes back with a result. It keeps long-term memory about you, about past conversations and about how it solved similar problems before. And it has time of its own: it can work on a schedule while you're not talking to it.

What your agent can do

  • Reply to text and voice notes, and read forwarded messages, photos and documents.
  • Search the web, read pages, compare options and send short summaries.
  • Run code on its own machine: write and test a script, clean up a spreadsheet, convert files.
  • Do things on a schedule: a morning brief, a Friday report, a nightly price check.
  • Save what worked as a skill and reuse it next time.
  • Find what you talked about weeks ago by searching its full conversation history.

A self-hosted AI agent in Telegram, minus the hosting

"Your own" is not a metaphor here. It stands for three technical facts, and together they are what control over your data actually looks like.

Its own machine

Each agent gets a separate, isolated Ubuntu 24.04 environment with a persistent disk. In the beta it's an isolated container on our server: no admin rights, no access to other users' machines or to our internal network. Everything the agent downloads, writes or saves stays on that disk across conversations and restarts. Dedicated virtual machines come later.

Its own memory: an AI agent with memory you can read

Hermes Agent keeps its memory in plain files on the machine's disk, not inside the model. MEMORY.md holds the agent's notes about its work and environment; USER.md holds what it has learned about you: preferences, how you like to be spoken to, dates that matter. Conversation history sits in a SQLite database with full-text search, which is how the agent finds something you mentioned a month ago. Skills are a folder of instructions the agent writes and improves on its own. You can view, edit, export and delete all of it in the Mini App.

Your own bot

Your agent lives in a Telegram bot of its own, not inside a shared service bot. The bot is created through Telegram's managed bots mechanism and belongs to your account; you choose its name and avatar. To Telegram it's an ordinary bot in your chat list, right next to people, and only you can talk to it. Next to the chat there's a Mini App that shows the machine's status, memory, skills and schedules. It opens in a regular browser too.

What still leaves the machine

To answer, the agent sends the text of a request to a language model: GLM by Z.ai via OpenRouter, routed to providers that don't keep prompts for training. The agent's web searches go to DuckDuckGo, the text of its voice replies is synthesized by Microsoft Edge TTS, and your voice notes are transcribed on the agent's own machine. Messages to the bot pass through Telegram's servers like any bot messages, and bot chats are not end-to-end encrypted. The machine runs in a data centre svoy rents. We say this plainly because "total privacy" from a cloud service would be a lie.

Managed Hermes Agent hosting

What Hermes Agent is

Hermes Agent is an open-source AI agent by Nous Research, built around a closed learning loop: after a complex task it writes a skill, improves that skill the next time it's used, nudges itself to save what matters, and builds a steadily better picture of the person it works for. It connects to Telegram, Discord, Slack, WhatsApp and Signal through a single gateway, transcribes voice notes, works with dozens of models, and has a built-in scheduler for tasks described in plain language.

Self-hosting Hermes Agent on a VPS

If you want everything on hardware you control, Hermes runs on a VPS with Ubuntu, no GPU required. The steps: install it with the one-line script from the official repository, run the hermes setup wizard, create a bot with BotFather, add the bot token and a model key, start hermes gateway and make it survive reboots. After that, updates, monitoring, backups and paying for keys are yours. For some people that's a pleasant evening. For others it's the reason they never start.

svoy vs your own VPS vs a marketplace image

There are three ways to get an agent. Your own VPS gives you full control and needs looking after. Ready-made images from cloud providers install Hermes for you, but you usually still configure the server, the updates, the model keys and the bot. svoy is the managed option: the machine, the model, the bot and the updates are included, and inside it's the same open-source Hermes Agent. If you ever move to your own server, your memory and skills move with you: "Download everything" gives you an archive that runs on plain Hermes.

Model and pricing

The agent thinks with GLM by Z.ai, through OpenRouter. We picked it for how it handles tools and long tasks, and for a sensible price. The model is included: no provider accounts, no watching API balances. Right now svoy is in beta: free, with a monthly model budget.

Who it's for

People who already live in Telegram and want their assistant there too, not in yet another app. People who care where their data sits. Developers who could run an agent themselves but don't want to babysit a server. And people who have never opened a terminal but have wanted "an AI of their own" for a while.

Limits

The agent makes mistakes, like any model, so check anything important: money, health, contracts. It works only on its own machine and gets no access to your accounts until you connect them yourself. The beta has no automatic machine backups — "Download everything" is your backup. Right now svoy works in Telegram and in the browser; other messengers and compute platforms will come later.

Questions

What is svoy?

svoy is a service that creates a personal AI agent in Telegram with one button. The agent gets its own isolated Ubuntu machine with a persistent disk, its own memory and its own Telegram bot. Under the hood is the open-source Hermes Agent by Nous Research.

How is this different from ChatGPT and other AI chats?

A chat answers questions inside one conversation. An agent remembers you across conversations, uses tools on its own computer (a terminal, files, web search), works on a schedule and can message you first. And it lives in your Telegram, not in a separate app.

Do I need a server, API keys or BotFather?

No. svoy sets up the machine, the model and the bot: you pick your agent's name, avatar and character, and confirm the bot creation in Telegram's window. No keys, no terminal, no server settings.

Where is my agent's memory stored, and what happens to it if I leave?

Memory is files on your machine's disk: MEMORY.md, USER.md, the skills folder and the conversation database. You can view, edit and export them in the Mini App. If you delete your agent, the machine and memory are erased right away — download everything first; the archive runs on any server with Hermes Agent. The bot stays yours.

Who can read my conversations?

Request text is processed by GLM by Z.ai via OpenRouter, the agent's web searches go to DuckDuckGo, and messages to the bot pass through Telegram's servers; bot chats are not end-to-end encrypted. The machine runs in a data centre svoy rents. We don't read your conversations or use them to train models.

Which model does it use, and can it change?

Today it's GLM by Z.ai via OpenRouter, included in the service. Hermes Agent works with dozens of models, and memory and skills live on the disk, not inside the model, so switching models later won't touch them.

Is svoy just Hermes Agent? Are you Nous Research?

svoy runs Hermes Agent, an open-source project by Nous Research under the MIT license. svoy is independent and not affiliated with Nous Research; we deploy, update and look after Hermes Agent for you.

Can I use it without Telegram?

To talk to the agent you need Telegram, because it lives in your bot. The Mini App, with memory, skills, schedules and machine status, also opens in a regular browser. Other messengers will come later.