How it works - One product, three things worth understanding.

Run Laminae as a managed per-tenant deployment, or take the same Docker Compose stack and run it on your own infrastructure. The wedge is open-standard, so the choice never locks you in.

Air-gapped works the same way the connected version does — bundled Ollama replaces the cloud model endpoints, and nothing leaves your network. There's no separate edition; it's the same code path.

Buckets are managed, vectorised knowledge stores. MCP servers are how your AI client talks to them. Non-technical teams can spin up both in an afternoon — no data team required.

Any MCP-compatible client connects: Claude Desktop, Cursor, Continue, an internal Copilot. Laminae returns evidence — the AI client generates the answer. You bring your own AI.

Dense plus sparse plus re-ranked. Every chunk is returned with its source document, page, and snippet, so the AI client — and the human reading along — can always verify the answer.

There is no narrative-generation step in the middle. Laminae returns evidence; the AI client decides what to do with it. That's why source attribution actually means something here.