AI note-taking that reads your own notes — not someone else's model of them

The pitch is usually the model: this app has AI in it. But a language model is a commodity you can rent by the token, and the part that is not a commodity is your notes. An AI feature is worth something only when it can read the specific, private, accumulated context that nobody else has — and that means the value lives in the vault, not in the assistant.

This inverts how the product should be built. Rather than a company holding your notes and selling you access to a model, you should hold your notes and bring whichever model you like. When a better model ships next quarter, you change a setting rather than a vendor.

Reading a single open note is table stakes and not very useful. The interesting operations need the whole graph:

Nodum's assistant runs on a key you provide. Claude, OpenAI, Gemini and Qwen are all supported, keys are encrypted at rest, and a key can be set per vault — so a work vault and a personal vault need not share a provider or a bill.

There is no per-seat AI upsell, because there is nothing to upsell: the software is MIT licensed and the model is yours. If you self-host and point it at a local model endpoint, no note ever leaves your network.

The assistant in the app is only one route in. Nodum is also a Model Context Protocol server at /api/v1/mcp — 36 tools over the same services and the same ownership checks the app uses. Point Claude Code, Claude Desktop or Cursor at it with a per-user token and the AI you already work in can create vaults, write and link notes, colour folders, search, import and export.

Tokens are per-user, hashed, and revocable without touching your session, so handing an agent access is not the same as handing it your password. Long-running operations like an import report progress as they go.

This is the part that makes an "AI brain" more than a phrase: the notes are a real store your other tools can read and write, not a chat window with a memory feature.