What Harbor is

Your team decides things all day: in pull request reviews, in Slack, in a doc nobody re-reads, and in the sessions your engineers run with their coding agents. Harbor reads the tools you already use, turns what was settled into short rules, and hands those rules to your agents at the moment they start working. Then it counts which ones the agents were given and which ones they said they used, so a rule nobody uses can be retired instead of piling up.

On the default setting a routine learning goes live straight away, on trial, with the sentence it came from. What only a person can answer waits for one in Review, for example a secret, an install step or a risky command, an author Harbor cannot match to anyone in your workspace, or a change to a live rule that the person behind it cannot make on their own. The rules on what waits are the same on every surface below, and what a coding agent remembers, and who approves it lists them.

Which of these is yours

I write code in a terminal.

Start with harbor context. It needs no account, no token and no network, and it prices what your agents already carry on every turn.

CLI →

I use Claude or ChatGPT and install nothing.

Connect Harbor as an MCP server and your assistant can search and cite your team's rules in the app you already have open.

Agents and MCP →

I am setting this up, for me or for a team.

Create your account, choose your Spaces, run one command in your repo, open your agent. A team adds GitHub, Space access and invites.

Quickstart →

Everything here

Four moves, and they are the loop in order. Something goes in, an agent uses it, you decide what stays, and you can see whether any of it worked.

Orient

what this is, and which path is yours

QuickstartGive Claude Code, Codex, Cursor and GitHub Copilot your team's memory: from sign-up to a working Harbor in one command, then three steps for a team.
Plans and limitsHarbor's Free plan limits, what you see when you reach one, what keeps working, and how to ask for Pro.

Wire

get your agents, and your team, reading and writing

harborloop CLIharborloop in full: every command and flag, which config files it changes, and what harbor doctor checks. The first command needs no account.
Agents and MCPAdd Harbor's MCP memory server to Claude Code, Codex, Cursor and GitHub Copilot: the four tools, and what the hooks add.
Slack tutorialSave a decision from Slack with @Harbor remember, ask with /harbor, and read every reply.
Ask from SlackA Slack knowledge bot that answers from your team's sources, cites every claim, and replies only to you.
GitHub App tutorialInstall the Harbor GitHub App, pick its repos, turn on Code review, and read the check, the comments and every @harbor-brain reply.
Code reviewCheck each pull request against your repo's rule files and your team's brain, with a comment on every line that breaks a rule.
Rules files comparedCLAUDE.md, AGENTS.md, Cursor rules and Copilot instructions: which agent reads which file, and how Harbor imports yours for review.

Feed

decide what goes in, and who it is for

SourcesConnect GitHub, Slack, Notion, Jira and more: what each connector reads, what it never reads, and how to scope it.
Knowledge and reviewWhat coding agents remember from sessions, pull requests and Slack, what waits for a person, and how unused rules leave.
Spaces and accessScope agent memory by repo and team: Spaces, roles, and who can read, approve and connect sources.

Prove

find out whether any of it worked, and what your agents may not do

Measure rule useWhether Claude Code, Codex and Cursor follow your team's rules: what each agent was sent, what it cited, and why that count is an upper bound.
GuardrailsBlock git push --force, rm -rf and database resets with hook guardrails (PreToolUse and its equivalents) in Claude Code, Codex, Cursor and GitHub Copilot.

What Harbor sends, what it leaves alone and how to remove it is on the home page, and the Privacy Policy has the full table.

How do coding agents share memory across a team?

On their own, they mostly do not. An agent's memory is tied to one repo or one machine: a CLAUDE.md or AGENTS.md file in the repo, Cursor rules, Copilot instructions, and whatever the agent saved in one engineer's session. A rule settled in a review on one repo never reaches the agent working in the next one, and a correction one engineer gives their agent is gone by the next session.

Harbor is the shared part. One command, npx harborloop init, wires Claude Code, Codex, Cursor and GitHub Copilot to an MCP server and to hooks on each machine. A rule learned in any engineer's session, a pull request review or a Slack thread, once it is live, reaches every engineer's agent from their next session, scoped to the repos it belongs to. The quickstart for team memory in Claude Code and Codex takes you through it, and scoping agent memory by repo and team covers who gets which rules.

What does Harbor add to CLAUDE.md and AGENTS.md?

A rules file is the right place for what an agent must read in one repo, and Harbor does not replace it. When you turn Harbor on in a repo, it offers the rules files already committed there (CLAUDE.md, AGENTS.md, Cursor rules and copilot-instructions.md) to Review, so they become team rules every agent can recall, in every repo they apply to.

What a file cannot do on its own, Harbor adds: it learns what the team decides outside the file, it hands an agent only the rules that apply to the task instead of the whole file, it shows which rules no agent uses so they can be retired, and it can stop a command your team forbids instead of hoping the agent read the line. See CLAUDE.md vs AGENTS.md vs Cursor rules vs Copilot instructions for which agent reads which file, and guardrails that block risky agent commands with hooks for the rules that must hold.