Ten minutes to your first served rule.

The six steps

1. Sign in and name your workspace

   Once you have access, signing in the first time asks for one thing: a workspace name. That is the boundary everything else sits inside, so use the company or the team, not the project.

2. Connect GitHub, and nothing else yet

   Go to Sources and connect GitHub. Merged pull request reviews are the richest source of a team’s real conventions: they are arguments that were settled, by people who had to live with the outcome.

   Resist connecting everything at once. A first review queue fed by six sources is a queue nobody reads, and the whole product depends on somebody reading it. You can add the rest once you have approved a few.

3. Install the CLI in one repository

   In a repository your team actually works in, not a scratch clone:

   bashCopy

   npx harborloop init

   This signs the machine in, installs hooks for every agent it finds, binds the repository to a project, and schedules the background pass that keeps it current. It is the step that makes everything after it measurable.

4. Work normally

   Open your agent and do a real task. There is nothing to type and no command to remember: knowledge arrives at the start of the session, and Harbor prints a short receipt into your transcript saying what it served.

   The first session may be served nothing, because the import from step 2 is still running. That is expected. Come back to it.

5. Approve one proposal, and edit it first

   Open Review. Everything an agent captured, and everything the import proposed, is waiting there staged. Nothing has reached your team yet.

   Pick one and edit it before you approve it. This is worth doing deliberately on the first one: it is where you find out that the thing you are approving is a sentence you would be happy to see in front of an agent tomorrow, and that you are allowed to change it.

6. Run harbor doctor

   Back in the repository:

   bashCopy

   harbor doctor

   The top half is configuration. The bottom half is the loop, and it is the part worth reading: how many sessions there were, how many received anything, how much of what was served went uncited, and what this repository has contributed back.

   If Attribution is not 100%, fix that before you read anything else on the page. It means sessions are not being counted, which makes every other number smaller than the truth.

What you just built

A loop. Something went in from a source your team already uses, a person approved it, an agent was served it without asking, and a number now says whether the agent used it. Every page in these docs is one part of that sentence in more detail.

Two things to do next. Read the CLI page for harbor sync, which puts your team’s skills and documents on disk, and for harbor off, which is how you keep a throwaway session out of the graph. Then served versus cited, which is what the bottom half of doctor is actually telling you.