Skip to content

Local-first & privacy

CoSci is local-first by design. This is the wedge cloud AI-for-science structurally can’t copy: a PI’s unpublished data, grant ideas and competitive hypotheses are exactly what they will not paste into someone else’s cloud.

Where your data lives

  • Everything runs on your machine. The backend stores state in a local SQLite database under backend/data/ (disposable — delete cockpit.db to reset).
  • There is no CoSci account and no managed inference. You bring your own key, or run fully offline.
  • In online mode, data leaves your machine only when you act — and only to the AI provider or scientific tool you configured, for that call.

Versioned and audited

Every action is versioned, and every model/tool call is recorded in the Agents / Runs log. Combined with human-in-the-loop approval gates, that means:

  • You can see exactly what ran, with which provider, and whether it was cached or hit the network.
  • Sensitive actions wait for your approval — nothing consequential happens silently.

Why it’s a feature and a cost model

Local-first isn’t only a privacy story. Free users run on their own hardware and their own API keys, so the open-source core costs almost nothing to run at any scale — and the same “data never leaves” property is what privacy- and IP-sensitive settings (industry R&D, clinical, pre-publication work) need most.