For engineering teams using coding agentsFor teams using coding agents
Harbor is shared memory for your team’s coding agents. What one developer teaches Claude Code, a teammate’s Codex or Cursor can use on the next task it fits.
$npx harborloop initno. exports are for paid plans only
Exports are for paid plans only.
CursorSamFrom Dana
+ if (!account.isPaid) return showUpgrade();CodexPriyaFrom Dana
+ if (!account.isPaid) return showUpgrade();CopilotLeeFrom Dana
+ if (!account.isPaid) return showUpgrade();Built-in memory stays inside one tool, and a rules file holds only what somebody stopped to write down. So the same lesson gets taught again, by the next developer, in the next agent.
DanaClaude Code, Monday
> add CSV export to reports
+ router.get('/reports/export', exportCsv);
> no. exports are for paid plans only
Stays in Dana’s Claude Code.
PriyaCodex, Thursday
> add a CSV download to invoices
+ router.get('/invoices/export', exportCsv);
> no. exports are for paid plans only
Priya teaches it again.
DanaClaude Code, Monday
> add CSV export to reports
+ router.get('/reports/export', exportCsv);
> no. exports are for paid plans only
Exports are for paid plans only.
PriyaCodex, Thursday
> add a CSV download to invoices
+ if (!account.isPaid) return showUpgrade();
+ router.get('/invoices/export', exportCsv);
⚓ Harbor your agent used 1 team rule
#1 Exports are for paid plans only.
Priya’s agent used Dana’s lesson.
Dana and Priya share a Space, and each ran one command. On the default setting a lesson goes live on trial and reaches the next session it fits. You can make lessons wait for approval. How Spaces work
Start freeA controlled test, not a customer result
Your agent can read your code. It can’t read what your team decided. So we gave it the same tasks with and without that one rule, and measured.
| Without the rule | With Harbor | |
|---|---|---|
| Right on the first try | 0 of 10 | 10 of 10 |
| Corrections a person typed | 10 | 0 |
| Agent time, mean | 52s | 32s38% less |
| Token bill per run, mean | 12¢ | 13¢About even |
Where it changes little. On two more tasks the code already held the answer, so the agent got them right either way: one ran faster, one a little slower. All four together: 26% less agent time.
Why the bill is even. The token count fell 34%, but most of that was cached reads, which cost little. Harbor’s tools and rules add some back, and the redo it saves pays for them.
What it does not show. Claude Sonnet 5.5 on a small Node repo, tasks we picked, one rule already saved, five runs per task each way. We have not measured what a team gains over time. Your numbers will differ.
No new tool, and no rules to write. Harbor learns from the work your team already does, and each developer’s agent starts with what fits the task.
It learns from your agents
And your team, if you connect it
Harbor
Exports are for paid plans only.
You, in Claude Code
Live, on trialWhat agents learn comes back
Each agent gets what fits, before it starts
Anyone can ask it
/harbor in SlackEvery PR is checked, if you turn it on
Code review
The knowledge your agents start with checks the pull request too. Harbor reads what your team has learned, plus your AGENTS.md and CLAUDE.md, and comments in GitHub on the line that breaks a rule.
Harbor
Exports are for paid plans only.
You, in Claude Code
Reaches each agent when it fits
Every PR is checked against it
Add CSV export to the free dashboard #5
Examplesrc/exports/export-csv.ts
13 function exportCsv(account) {14- requirePlan(account, 'paid');14+ requirePlan(account, 'free');harbor-brainbot
Team rule · src/exports/export-csv.ts:14
requirePlan(account, 'free') never throws, so a free-plan account calling exportCsv now gets the full CSV. The team rule says exports are paid-only. Revert to 'paid', or change the rule first.
Fix: Revert to requirePlan(account, 'paid'), or get the team to change the exports rule first.
Rule: Exports are for paid plans only. They are what sales sells the upgrade on.
Next PRHarbor review 41 rules checked, from AGENTS.md, CLAUDE.md and the team’s brain · none broken
Sign up, connect GitHub, turn on Code review. It works beside your tests, CI and reviewers and replaces none of them. A broken rule gets a comment and never blocks the merge. How it works
Start freeA rules file holds what somebody remembered to write, and every session reads all of it. Harbor learns from the work, reaches every developer’s agent, and each task gets only what fits. Keep your files: Harbor sits beside them.
>no. exports are for paid plans only
⚓ Harbor saved a ruleYour team’s agents can get it from their next session.
Outside research, not a test of Harbor
Over 138 tasks, a hand-written context file moved task success 2.4 points, not a significant gain, and cost about 20% more to run. ETH Zürich and LogicStar, arXiv:2602.11988
No call, no rules to write and nothing committed to your repo. Keep the agents you have.
Google or email, then one field. No card.
Setup runs in a terminal, so we email you the sign-up link to open on your computer.
It signs you in through the browser and wires each agent it finds. If the repo has rules files, it asks before sending them to review.
$npx harborloop init
Open a new session and work as you do. Your first correction is the first lesson.
Invite them to the repo’s Space from the app. They run the same command, and their agent can start using what yours learned.
$ npx harborloop init
Opening https://app.gethrbr.com/desktop-auth
Waiting for the browser…
Harbor is set up. Workspace "Acme".
claude ~/.claude.json
codex ~/.codex/config.toml
cursor ~/.cursor/mcp.json
this repo acme/app is on, Space "General"
Next: open a new Claude Code session in
this repo.Working alone? A lesson still follows you to your other agents, repos and machines. Free does not expire, and reaching a limit never stops your agents getting their rules. See the limits
It is off until you turn it on in a repo. Here is what it sends, what it leaves alone, what it changes on your machine and how to get rid of it.
.env, .pem and other secret files.~/.harbor.harbor off pauses it: no context in, nothing recorded.harbor uninstall removes it from the machine.The text is processed by OpenAI and Anthropic. Harbor runs on AWS in us-east-1, encrypted at rest. There is no SOC 2 report yet. Connected sources are optional, and each runs under a credential you issue and can revoke.
Code review is off until an admin turns it on. Each review clones the repo in a short-lived task, sends the changed code and the parts it reads to Anthropic, and keeps no code. A public repo is checked against its own rule files, never your team’s rules.
Read the Privacy PolicyWhat each source readsWhat Code review reads
If you only use Claude Code, in one repo, on one machine, its auto memory covers a lot. It stays in that repo on that machine, only Claude Code reads it, and the first 200 lines or 25KB load into every session. Harbor reaches your teammates and their agents, plus your other repos and machines, gives each task only what fits, and shows you which rules an answer cited. It is free to try, and harbor uninstall takes it back out. How auto memory works.
You correct Claude Code in a repo that is on, in words that state something standing, and Harbor saves it to that repo’s Space. Anyone in that Space who has run npx harborloop init can get it from their next session, in Claude Code, Cursor, Codex or Copilot, when it fits the task. On the default setting it goes live on trial with no approval, and you can make lessons wait for a person. One-off fixes and edits you make by hand are not saved. How Spaces work.
Free costs nothing for you and one teammate. There is no card and no end date, and you only pay if you ask for a paid plan. It has limits: one source, 1,000 live items, 30 days of history, and 200 correction reads and 10 questions a day. Reach one and your agents keep getting their rules. Every limit, and what happens at it.
No. On your own, a correction follows you to your other repos, agents and machines, and you can read and edit everything Harbor keeps. Add teammates when you want them, and Free covers you and one more.
No. Harbor reads your corrections as you work, and your agent saves what cost it time when a task ends. Start a line with remember: when you want to be sure. If your repo already commits an AGENTS.md or CLAUDE.md, setup asks before sending its rules to review, so nothing you wrote is left behind.
Press Start free and sign up with Google or email. Name your workspace, then run npx harborloop init in your repo. It signs you in through the browser, finds Claude Code, Cursor, Codex and GitHub Copilot on your machine and wires each one with hooks and the MCP server. There is no call and no card. harbor doctor tells you if something is not working. Read the quickstart.
Connect them to Harbor. One setup wires Claude Code, Cursor, Codex and GitHub Copilot to the same memory, so a lesson saved from one can reach the others from their next session. Harbor learns from the work, sets a gate your team controls, gives each session only the rules that fit, and counts which rules got cited. Shared memory for AI coding agents.
Keep them. They hold what somebody remembered to write, and each tool loads them its own way. Harbor sits beside them: it learns from your team’s sessions, reaches every agent your team uses, and hands over only what fits the task. It never edits your files. How it works with rules files.
No. It checks your team’s rules and nothing else: your repo’s rule files and what Harbor learned from your work, the same rules your coding agents get. It does not replace tests, CI, security scanners or your reviewers. Other reviewers can follow instructions you write for them, and some learn from their own reviews. A broken rule gets a comment on its line and never fails the check. It can miss one, and it stays quiet when it cannot confirm it. How the review works.
Yes. A very short correction, or one its model does not read as a rule, is not saved, and nothing is captured outside a repo you turned on. Start the line with remember: to be sure, or add the rule yourself in Knowledge.
On the default setting, a rule someone saves on purpose replaces one it reverses when they could have retired the old one themselves, and the old one keeps a Restore button in Knowledge. Otherwise, and whenever a correction reverses a rule, the pair waits for a person. When a push deletes a file a rule names, Harbor retires that rule on its own if it belongs to one repo, and flags it as out of date if it reaches more. And a rule agents are given but never cite is marked in Insights, so you can retire it with evidence. What Harbor counts.
Not yet. We measured one controlled test with one agent and one saved rule, and it is on this page with what it leaves out. We have not published a result from a team over time. The way to find out is cheap: Free covers two people, so run it on a real repo with one teammate and read which rules got served and cited. See the test.
Every day your company learns more about how it works. Its agents should benefit. We’re starting with engineering.
Free for you and one teammate. No card. No end date.