AI on your terms
At Palimpseste, connecting artificial intelligence is not a technical setting: it is a matter of sovereignty. Three stances, from the most closed to the most open — and you can switch between them whenever you like.
IOn your machine
A model installed on your own computer, with Ollama or LM Studio. Nothing leaves the machine — not your texts, not your requests, not the replies. It works on the train, in the library, in the field; it costs nothing to run; and it is the natural stance for sensitive data. The setup guide gets you up and running in half an hour.
IIOn your network
The server in your lab, your newsroom or your firm — a shared GPU behind an OpenAI-compatible gateway (LiteLLM, vLLM), reached directly, through an SSH tunnel or over WireGuard. Sovereignty for the team: the infrastructure you already own, confidentiality you control, computing power you share. Our apps connect to it as they would to any provider: one address, and that is all.
IIIIn the public cloud
The best models of the moment — Anthropic, OpenAI — with your own key. You pay the provider, at the provider's price: no credit-based subscription, no middleman reselling tokens. And for a corpus that must not be exposed, the range has its answer: Palimpseste Pseudo pseudonymises the folder before it leaves, and restores the identities when it comes back.
What each app does with it
| App | Stances | What you should know |
|---|---|---|
| The Editor | All three — and AI remains optional | Automatic discovery of installed models, silent repair of local-model failures, guaranteed output language, and the estimated cost is shown before every operation. The whole workshop also works without any AI at all. |
| Pseudo | Machine or network for its assistant; the public cloud is the airlock's destination | Automatic name detection relies on your local assistant (LM Studio, Ollama) — without one, you write the list yourself. If the declared address leaves the local loop, the app warns you; it does not block — you remain the operator responsible for your own network. And it is because the folder has been pseudonymised that it can then go out to public AI. |
| Loupe | None — and that is a stance in its own right | Loupe closes the axis: the app declares no network access entitlement at all. Its only models (HD magnification, point tracking) run on the device, through Apple's frameworks — and the app always tells you when a result is reconstructed rather than observed. |
Our safeguards
- AI is never a prerequisite for writing, restoring or looking. It adds to the work; it never gates it.
- What leaves, and where it goes, is always stated before anything leaves — never a silent request to a provider.
- Your keys are yours — stored in the system keychain, never with us.
- No magic compliance for sale: pseudonymisation helps, the review remains yours — our product pages say so in plain words.