Bring your own AI

Harbor has no AI built in. No assistant in the sidebar, no sparkle icon, no “smart” anything.

That’s not a gap we’re racing to fill. It’s one of the most deliberate decisions in the product, and it deserves a full explanation.

Why we didn’t build one

You almost certainly already pay for an AI. ChatGPT, Claude, or one of their peers, models built by companies spending billions to make them good. Any assistant a notes company bundles will be worse than the one in your other tab, and it will come with two costs you didn’t ask for.

The first cost is money. In this category we’ve watched a bundled AI become the excuse for a price hike: a feature nobody requested, used to justify charging everyone more. We’re not doing that, and the cleanest way to not do it is to not have the feature.

The second cost is trust. A built-in assistant means the company’s AI is reading your notes, on their terms, under a policy that can change after you’ve moved your life in. Your second brain shouldn’t come with a mandatory reader.

What we built instead

We made Harbor open enough that the AI you already trust can work your notes, with you holding the keys:

• A documented REST API. The whole product is reachable over an open, documented API, with a full OpenAPI spec any tool can consume, secured by personal access tokens you create and revoke yourself.
• The harbor CLI. Every part of the API, exposed as clean commands with --json output on all of them. Built for scripts, automations, and the way modern AI agents actually like to work.
• An agent skill. One command, harbor skill install, teaches Claude Code, Codex, or Cursor how to use Harbor properly: creating notes, formatting them richly, organizing notebooks, setting reminders, searching your library. Your AI becomes a genuinely capable notes assistant in about a minute.
• An MCP server, at launch. For every tool that speaks the Model Context Protocol, Harbor will be a first-class connection from day one.

What does this look like in practice? You ask the assistant you already use to file today’s meeting notes into the right notebook and tag them. To dig up the invoice from the fence project and summarize the payment terms. To build a packing list note from your last three trip notes. Your AI does the reading and writing; Harbor is the place it reads from and writes to.

The AI is as good as the one you chose, because it is the one you chose.

Your keys, your control

Nothing is connected until you connect it. Access runs through personal tokens you issue, scoped to you, revocable anytime in settings. Unplug your AI and Harbor goes quiet again; it works completely without one.

And your notes never train anyone’s model through us. We have no model to train, and we don’t hand your data to anyone who does. What you choose to send to your own AI provider is governed by your account with them, under your control. That’s the point of the whole arrangement: you pick the AI, you set its terms, you hold the keys.

The encrypted exception

One boundary worth repeating, because it’s the same one everywhere in Harbor: encrypted notes are unreachable through the API, the CLI, and MCP, just as they’re excluded from OCR and search. We can’t read them, so nothing you connect can either. Lock down the sensitive things, and your AI works with everything else.

Openness is the feature

If a note-taking company shows you a shiny assistant, ask what happens when it’s worse than the AI you already have, and what it costs you to be its captive audience. We’d rather compete on being the best harbor for your notes, and let the best AI in the world, whichever one that is this year, be the one you bring.

The full picture, including examples, is at bring your own AI. And you can start free at Harbor and plug in your own assistant whenever you’re ready.