What leaves
your desk.
Short answer: as little as possible, and never the things that make you you.
inteligentia is local-first, not air-gapped. Your agent lives on the box and does most of its work there. When a task truly needs a frontier model, the box sends a filtered request to the provider you chose. Here is every step.
The path of one request
You message your agent on Telegram or another chat app. The box on your desk picks it up. We are not in the middle.
The on-device model and the index of your files try to answer. Most daily questions end here.
If it needs more horsepower, the filter masks names, numbers, addresses and IDs, and drops context the task doesn't need.
The thin, masked slice goes to your provider (OpenAI Codex or OpenRouter) on your own account.
The answer returns, the real details are put back in locally, and the conversation is saved on the box.
On the box
- Your agent and everything it remembers about you
- Every chat, past and present, including the history you import from ChatGPT or Claude
- Your files and the search index built from them
- The real names, numbers and addresses the filter masked
- The answers you got back from the cloud, once they arrive
To your provider
- The masked text of a request that needs a frontier model
- Only the context that task needs (the clause in question, not the whole contract)
- Nothing goes to inteligentia. We are not in the middle.
Want zero exposure for a while? Disconnect the cloud brain. The agent keeps working on the box, just with a smaller brain.
What we won't pretend
- It is not 100% private. When the agent asks a cloud model for help, filtered text leaves your desk. That's the trade for frontier-level answers, and it's your choice.
- The filter is automatic, not magic. It catches the usual suspects (names, IDs, account numbers, phone numbers, emails, addresses). Unusual details written in free text can slip through, so keep anything truly secret in local-only conversations.
- Once text reaches a provider, their policy applies. Check your provider's data settings. With OpenRouter, the request is served by the model host you pick.
Built on open source
The agent at the heart of the box is built on Hermes Agent, an open-source (MIT) project by Nous Research. We package, tune and look after it so you never touch a terminal. The core stays inspectable, and you're never locked in to us either.
We practise it here too
No analytics, no trackers, no cookies. The homepage demos run in your browser. The only thing this site ever sends is the reservation form, and it shows you the exact payload before you press the button.
Sources for the comparison
The "Rented intelligence" table in the homepage preview summarises each vendor's public pages as read on 1 Nov 2026. "Not documented" means we found no published way to export the agent's memory or run it on another provider's model. Products change fast; if something here is out of date, tell us at info@inteligentia.eu and we'll fix it.
| OpenAI dots | openai.com/index/introducing-dots · help.openai.com |
|---|---|
| xAI Grok Bot | x.ai/news/introducing-grok-bot · x.ai/news/designing-grok-bot |
| Anthropic Claude Duet | claude.com/resources/articles/claude-duet |
| Meta Muse | theneuron.ai: How to get started with Meta Muse |