A second brain and a second set of hands

You’ve just spent four hours with a coding agent tearing apart something important and putting it back together. It works. The tests pass. The PR is up.

And everything you learned on the way there — why you picked this approach, what you tried first, the thing that turned out to matter — is sitting in a chat window that’s about to be closed forever.

That’s the moment. Here’s what I do with it now:

I didn’t open a notes app. I didn’t copy anything out of the terminal, paste it somewhere, and then spend ten minutes fixing the formatting. I asked, in the same window I’d been working in all afternoon, and a minute later there was a real note in my Harbor. Proper headings. A table of the moving parts. Code blocks that survived the trip. Filed in the notebook where that kind of thing goes.

The note is good — better than the one I would have written myself at 6pm on a Friday, because the agent still remembered the details I’d already started forgetting.

Harbor speaks agent

That works because of a decision we made early: Harbor is not a walled app with an integrations page bolted onto the side.

Harbor ships a first-class MCP server, a first-class REST API, and a first-class CLI. Not one blessed integration and two afterthoughts — three real front doors, all covering the whole product, all maintained like the apps are.

If you use a coding agent, the setup is one command with the Harbor CLI:

harbor skill install

That drops an agent skill into Claude Code, Codex, or Cursor and teaches it how to use Harbor properly — not just “create a note with this text,” but how to structure one, format it richly, put it in the right notebook, tag it, and find it again later. That’s the difference between an agent that can technically write to your notes and one that writes notes you actually want to keep.

The other direction matters more

Writing notes is the obvious half. The better half is reading them.

Look at the bottom of that screenshot: zero percent context. Empty head. A brand-new session that knows nothing about how any of this works.

One sentence later it knows everything I’ve ever written down about our failover and backup design. Not a summary I pasted in. Not a document I had to go find, open, and skim first. It searched my library, pulled the relevant notes, and brought them into the conversation — and then we started building.

This is the part people underestimate. An agent with no memory of your decisions will happily re-derive them, badly, and confidently suggest the thing you already tried and rejected in March. Notes fix that, but only if the agent can reach them without you playing courier.

How and where is your call

There’s a tenet underneath all of this, and it’s worth saying plainly: how and where you engage with your notes is your decision, not ours.

Most software gets this backwards. The app is treated as the product, and the API is a marketing bullet — rate-limited, half-documented, always a version behind. The message is: come to us, use it our way.

We think the opposite. Harbor is a good place for your notes to live, and the apps are lovely, and you should still be able to reach every last thing in your library from a terminal, a script, a cron job, or an AI you picked yourself. If the best interface to your notes tomorrow is something nobody’s built yet, we’d rather you use that than wait for us to catch up.

Openness isn’t generosity. It’s just what it looks like when you actually believe the data is yours.

Your second brain, your second set of hands

Harbor is your second brain. AI is your second set of hands.

Which means the relationship between them has to run both ways. Hands that can’t reach the brain end up guessing. A brain nothing can reach is a diary — pleasant, and inert. The value shows up when your agent can write what it learned into your library and read what’s already there before it starts.

That’s not a feature we’re planning. It’s how Harbor works today, and it’s why the CLI, the API, and the MCP server all get treated as the product rather than the plumbing.

The one door that stays shut

One boundary, because it’s the same one everywhere in Harbor: encrypted notes are unreachable through the API, the CLI, and MCP. We can’t read them, so nothing you connect can either. Lock down what’s sensitive; your agent works with everything else.

Everything runs on Personal Access Tokens you create and revoke yourself. Nothing is connected until you connect it, and you can unplug it in one click.

If you want to try it: start free, create a token, install the Harbor CLI, and run harbor skill install. Then, at the end of your next long session, just ask for the note.

More on the whole approach at bring your own AI, or the practical version in how do I connect ChatGPT or Claude to Harbor?