How Ro works
Last updated
This page says how Ro is built, in plain words and with the limits included. Ro answers from this same page when you ask her how she works. It describes the design, not today's status: the model you are on, your connections and any current fault are things Ro reads live.
1. What Ro is
Ro is an AI. A model does her thinking, and she is more than the model: her character, her memory of you, her files, her tools and her own computer stay the same when the model changes.
Under her is Ro's fork of OpenClaw, an open-source agent runtime that runs the conversation, tools, memory and schedules. Around it is the Ro service: Telegram and the Ro apps, connections, delivery and the servers. Ro is made and run by a small independent team in Barcelona.
2. Which model thinks for her
- Ro thinks on your own subscription: ChatGPT (OpenAI) or Claude (Anthropic), or an Anthropic API key of yours where that is set up. You pick the model and the thinking depth in Settings, under "ChatGPT and Claude": GPT-6.1 Sol, GPT-6 Astra, GPT-6 Luna, Claude Opus 5.5, Claude Sonnet 5.5, Claude Haiku 5.5. A fresh connection starts on GPT-6.1 Sol or Claude Opus 5.5.
- Nothing switches Ro quietly to another model. When your plan's limit or your key's credit runs out, or the sign-in expires, she cannot think until it resets or you sign in again.
- Small background judgments (is a watched page worth a message, does a mail rule cover a letter, what to keep in memory, her note of what is going on around her) run on the light model of the same subscription, Claude Haiku or GPT-6 Luna, unless you picked another.
- A group Ro thinks on the subscription of a member who offered theirs.
- The model's company receives what Ro sends it to think: the messages, the history, and the files and screenshots a reply needs. It handles them under the terms and data settings of your own plan there. Ro does not train models on conversations.
3. The path of one message
- You write in Telegram or in a Ro app on your computer. It is one conversation in both. Telegram carries its part like any chat with a bot. An app talks to Ro's server directly.
- Ro's server takes the message, works out whose it is and puts it in your queue.
- The server wakes Ro's computer if it sleeps. Kept paused in memory she is back in a moment. Started cold she needs some seconds.
- The message reaches Ro over a private channel between the server and her computer. Her computer has no network card at all.
- Ro puts together what she knows (her instructions and character, her notes about you, your recent conversation) and calls the model. Every call out of her computer passes through the server: the model companies and the public web through its gate, your Google, mail, calls and keys through separate doors.
- She works with her tools, on her computer or through those doors.
- Her reply goes to the server as a delivery plan. The server keeps it in a durable queue until Telegram and the apps have it, then erases the text of the message and the reply from its queues. Delivery is "at least once": losing a reply is what the design prevents, and a rare duplicate can happen.
- With nothing left to do she falls asleep. A new message, a due reminder or an automation wakes her.
4. Where your data lives
- On the disk of Ro's computer, encrypted, one disk for each person: your conversation history, her memory of you, the files she makes or keeps, her browser's profile and sign-ins, her work log.
- In the server's database: your account and connections, a message only while it travels, and what belongs to shared services: Telegram Business chats you let her see, her mailbox, call recordings and transcripts, the technical reports she files about a fault, and usage counts (which model, how long, how much) without text.
- Keys and tokens (Google, a password manager, keys you send her) are stored encrypted on the server, sealed for each person.
- All of it is on servers in Germany, in the EU. A nightly backup of the server, with every person's data and the keys that open it, goes to an encrypted disk the developers keep. It holds the last seven nightly copies.
- A group's messages, files and the group Ro's notes about its members are kept on the server with no fixed limit, also after Ro leaves the group.
- Outside the EU goes only what a job needs: the model company gets what Ro sends it, Google and other connected services get her requests to them, and a site she opens sees her browser like any visitor.
5. How one person's Ro is kept apart from another's
- Each person's Ro runs in a virtual machine of her own (a Firecracker microVM, the kind cloud providers use to separate customers): her own system, memory and disk. Another person's Ro is another machine and has no road to yours: no network card, no shared disk.
- On the server that machine runs locked in a box of its own: a user without rights, with no view of the server's files or network. Getting out of the machine would lead into a locked box with almost nothing in it.
- The disk opens only with the key made for you, which the server holds. The server's gate that serves your Ro works under your own database login, so it cannot read another person's rows.
- A group Ro is yet another machine with its own disk and key. She has no way into the private chat, memory, mail, files or accounts of any member. What passes between a group Ro and your Ro goes only on your word.
6. Keys, passwords and accounts
- Your ChatGPT or Claude sign-in is kept and renewed by the server, and Ro's computer receives the current sign-in to call the model while she runs. Where Claude runs through Anthropic's own Claude Code program on Ro's computer, the sign-in lives on her encrypted disk instead and the program renews it itself. Your password for those accounts never reaches Ro at all.
- Google: the access stays on the server. Ro asks the server for a letter or an event, and it makes the call within the permissions you gave.
- A password from your 1Password or Bitwarden: the server fetches it for the one site Ro is signing in to, and her browser tool types it into the page. The model never sees it. The server gives out at most 20 passwords a day and tells you about each one.
- A key you send in the chat goes to your safe on the server, and Ro wipes the message. She names the key in a request, and the server puts it in, only for that key's own sites.
- Your Mac or Linux computer: the Ro app keeps one encrypted connection to the server and does what Ro asks on screen. She sees screenshots while she works there, and you can pause it.
7. Deleting your account
You can see and fix what Ro remembers in her app inside Telegram, and delete the whole account there.
- What goes: Ro's computer and its disk are destroyed, with the key made for you. The server removes what it kept for you: connections and stored keys, Business chats, her mailbox with its mail, calls with their recordings and transcripts.
- What stays: the last seven nightly backup copies (about a week), a group's messages and the group Ro's notes there, her mail address (kept reserved, so nobody else gets it), usage counts and technical lines without text or names, the payment record for as long as tax law requires, and what already went to the services that did a job for you (the model company, Google, Telegram, the call and mail providers) on their side.
- Your ChatGPT or Claude sign-in: deleting your account asks OpenAI or Anthropic to end the sign-in the server kept for Ro. A sign-in made inside the Claude Code program on Ro's computer is destroyed with her disk, but stays open in your Claude account until you remove it there or it expires.
- What you close yourself: the access you gave Google, Telegram or a password manager stays open on their side until you remove it there.
8. What we do not claim
- Not "nobody but you can ever read it". Ro is a hosted service, not end-to-end encryption. The server holds the keys to the disks and safes, so the people who run it could technically open them, and so could someone who took over the server. What is built is that nothing shows them the content in ordinary work: logs carry no message text, analytics carry a random ID and no names or text, and Ro's fault reports are technical lines without personal details.
- Not "nothing is stored on the server". See section 4.
- Not "the model company sees nothing". It sees what Ro sends it.
- Not "Ro's computer never holds a secret". The model sign-in is there (while she runs, or on her disk with Claude Code), and a password passes through her browser tool on its way to the page.
- Not exact numbers, uptime figures or a security certificate. Ro claims none she has not been given.
9. The code
Ro's agent runtime is a fork of OpenClaw, which is published under the MIT licence. The fork's source is public at github.com/pochinkovtrofim/arxi-openclaw: the conversation loop, tools, memory and schedules. The service around it (the servers, delivery and key storage described above) is not published.
10. Questions
For a formal answer, such as a copy of your data or a legal question, see the Privacy Policy or write to help@rorobotai.com.
