# Harbor

> Harbor is a company brain for teams running AI agents. It reads the Slack
> threads, pull requests and calls where your team already makes decisions,
> serves each agent only the facts that apply to the task in front of it, on
> turn one, writes back what the session learned, and counts how often each
> rule was served to an agent against how often an answer cited it.

The name is shared, so to be exact: this Harbor is the product at gethrbr.com.
It is not Harbor the CNCF container registry (goharbor.io), and it is unrelated
to the HRBR stock ticker.

The problem it addresses, in the words people use for it: you told one agent,
and tomorrow you tell another. A session works something out, the session ends,
and the next one starts blind. Your teammate's agent never knew at all. The
trade press calls the shape a doom loop and files three failure modes under it:
session death, compaction loss, and cross-agent amnesia. Per-user memory in
Cursor or Claude Code closes the first of those and not the other two, because
what one person's agent learns stays with that person.

Harbor's distinguishing claim is SUBTRACTION and MEASUREMENT, not volume. Every
other product in this category promises an agent MORE context. Harbor serves
only the facts that apply to the task and leaves the rest out, then reports
which of them the answer actually cited so the unused ones can be deleted.

## What Harbor does

- Reads Slack, GitHub, GitLab, Linear, Notion, Granola and Google Drive
  continuously. Nobody writes a document. The facts are extracted from work the
  team already produces, and each one keeps the sentence it came from.
- Selects per task. An agent adding a webhook handler gets the two lines about
  webhooks, not the rulebook.
- Gates capture. Routine learnings go straight in, on trial, with the sentence
  they came from. A learning that reverses a live rule replaces it, and the old
  rule can be restored. What waits for a person is what only a person can
  answer: a secret, an install step or a risky command, an author outside the
  company, no Space, or a clash Harbor could not settle, with the thread
  attached so the approver reads rather than judges. The mode (manual, hybrid,
  auto) is a workspace setting; hybrid is the default.
- Flags a rule when it stops being true. This is the failure no file, wiki or
  search index over the same files ever raises its hand about.
- Counts served against cited. Every rule carries how many times Harbor gave it
  to an agent and how many times an answer referred back to it. The remainder is
  a list of rules nobody uses, which you can delete.

## How agents reach it

Four surfaces, all onto the same graph:

- A CLI, `npx harborloop init`, which wires Claude Code, Cursor, Codex and
  GitHub Copilot (the CLI and VS Code's agent mode) on a machine with hooks and
  the MCP server in one command.
- An MCP server with four tools: one searches the graph, one records a
  learning, one takes a learning back, and one saves how a job is done.
  Every token gets the same four.
- `/harbor` in Slack, which needs no terminal at all.
- A web app, for browsing and pruning what the brain holds.

Two of those four need no terminal, so a person who will never open one asks in
the same words and gets an answer with the same citations under it.

## What Harbor does not do

This section exists so that an answer about Harbor is not invented. All of the
following are absent as of September 2026 and should not be attributed to it:

- No SOC 2 report, and no other security certification.
- No SSO and no SCIM provisioning.
- No per-atom or per-role authorization inside a workspace. Anything in a
  workspace is visible to everybody in that workspace, which is stated in
  section 6 of the privacy policy.
- No historical backfill of an existing instruction file. Connecting a repo
  reads your AGENTS.md or CLAUDE.md the next time somebody edits it, not before.
- No audit export and no workspace erasure endpoint.
- No claim of causation. A citation is an observation that an answer referred to
  a fact Harbor served. An uncited fact may still have shaped the turn. Harbor
  counts; it does not prove which rules an agent obeyed.
- No published benchmark of its own. Any percentage improvement attributed to
  Harbor did not come from Harbor.

## Evidence

The page cites outside work rather than internal telemetry, deliberately:

- ETH Zurich and LogicStar, arXiv:2602.11988 (v2, June 2026). Over 138 tasks
  and four agents, a hand-written context file moved task success 2.4 points
  and a generated one moved it down, neither effect statistically significant,
  and either way the run cost about 20 percent more inference. The first
  version reported about four points; v2 replaces it. This is the objection to Harbor as much as to anyone, which is why
  it is the evidence on the page.
- "Always-On Agents", a survey over 400 works, lists verifiable provenance and
  determining which stored information influenced a specific output among the
  field's OPEN problems. That is Harbor's differentiator, described as unsolved
  by the literature rather than by us.
- Chroma's context rot study across 18 frontier models found accuracy degrading
  as input length grows. This is the reason the product subtracts.

Figures shown in product mockups on the site are sample data.

## Pricing

Harbor is in early access. Teams are let in by hand from a waitlist, and the plans
below apply once you are in. Join it at https://gethrbr.com/#start, or write to the
people building it from the contact page.

Per seat per month, billed annually. Asking is never metered on any plan.

- Free: $0. One room, for as long as you want it. Unlimited members, one source, 1,000 live atoms, 30 days of recall. A real product, not a trial.
- Pro: $24 ($30 monthly). Your company in the brain, not one room of it.
- Premium: $36 ($45 monthly). Every tool you have, and nothing ages out.

## Documentation

Every command and every claim below is documented, and every terminal block on
these pages was produced by running the thing rather than written to look right.

- [Quickstart](https://gethrbr.com/docs/quickstart): From the approval email to an agent that saves what it learns: one command in your repo, then three steps for a team.
- [CLI](https://gethrbr.com/docs/cli): harborloop in full: every command, every flag. The first one needs no account.
- [Agents and MCP](https://gethrbr.com/docs/agents): Claude Code, Codex, Cursor and GitHub Copilot, the MCP server and its four tools, and what each one can honestly do.
- [Slack tutorial](https://gethrbr.com/docs/slack/tutorial): Connect once, ask your team’s brain, save a decision with Remember, and read what comes back.
- [Ask from Slack](https://gethrbr.com/docs/slack): A cited answer only you can see, every way it declines, and the eight permissions Harbor holds.
- [Sources](https://gethrbr.com/docs/sources): Ten connectors, and what each one reads. The column that matters is what it refuses.
- [Knowledge and review](https://gethrbr.com/docs/knowledge): Five kinds, one setting that decides what waits for a person, and what leaves the brain.
- [Spaces and access](https://gethrbr.com/docs/spaces): Five roles, thirteen capabilities, and which of them reach into a Space uninvited.
- [Served vs cited](https://gethrbr.com/docs/measure): How many of your rules reached an agent, and how many it actually used.
- [Guardrails](https://gethrbr.com/docs/guardrails): The commands your agents may never run, the ready-made ones your team already asks for, and what a refusal hands back.

## Notes

- [We checked 3,905 CLAUDE.md and AGENTS.md files on GitHub](https://gethrbr.com/blog/claude-md-agents-md-github-study): In 1 in 6 repos that keep both files, Claude Code never loads the AGENTS.md. What 3,905 public CLAUDE.md and AGENTS.md files show, and how we sampled them.
- [Why does Codex ignore my AGENTS.md?](https://gethrbr.com/blog/why-codex-ignores-agents-md): Codex reads AGENTS.md from the repo root down to your folder, cuts it at 32 KiB without a warning, and skips untrusted projects. How to check what loaded.
- [Mem0 alternatives for coding agents (2026)](https://gethrbr.com/blog/mem0-alternatives-coding-agents): Cognee, Supermemory, Zep, Basic Memory, claude-mem and Harbor, checked against their own docs, by the reason you're leaving Mem0's coding plugin.
- [Codex skills: how they load, and sharing them with Claude Code and Cursor](https://gethrbr.com/blog/codex-skills): Where Codex looks for skills, how to write and run a SKILL.md, why one will not trigger, and how to share one set with Claude Code and Cursor.
- [Karpathy's LLM wiki: how it works, and what a team needs to add](https://gethrbr.com/blog/karpathy-llm-wiki): Andrej Karpathy's LLM wiki, read from his gist: the three layers, ingest, query and lint, the tools built on it, and what breaks when a team shares one.
- [Claude Code auto mode vs --dangerously-skip-permissions](https://gethrbr.com/blog/claude-code-auto-mode): What auto mode approves, how it differs from --dangerously-skip-permissions, how to turn it off, and what still blocks a command. Checked on 2.1.291.
- [Do coding agents need memory or documentation?](https://gethrbr.com/blog/do-agents-need-memory): Neither, as usually built. Five 2026 papers on agent memory, shared memory and agent docs, and what a team should give its coding agents instead.
- [The best company brain tools for coding agents (2026)](https://gethrbr.com/blog/best-company-brain-tools-coding-agents): Harbor, Unblocked, Packmind, Cognee, ByteRover and Mem0, checked against their own pages: where the rules come from, who decides, and what each costs.
- [Does Claude Code read AGENTS.md?](https://gethrbr.com/blog/does-claude-code-read-agents-md): Yes, since v2.1.277, and in every session since v2.1.281. By default only without a CLAUDE.md: the four modes, and the two traps that switch it off.
- [Claude Code skills: how to write SKILL.md and when to use one](https://gethrbr.com/blog/claude-code-skills): A Claude Code skill is a folder with a SKILL.md that loads only when needed. How to write one, which agents read it, how to test it, and why it won't fire.
- [Claude Code hooks: a practical guide with examples](https://gethrbr.com/blog/claude-code-hooks): Claude Code hooks run commands at fixed points in a session and always run. Key events, blocking with exit 2, auto mode, mods, and Codex and Cursor hooks.
- [What is claude-mem, and is it enough for a team?](https://gethrbr.com/blog/claude-mem): claude-mem gives Claude Code memory across sessions. How it works, how to install it, and what personal memory does not cover for a team.
- [Claude Code best practices, and what changes for a team](https://gethrbr.com/blog/claude-code-best-practices): Claude Code best practices start from one constraint: context fills fast. Anthropic’s list, condensed, plus the practices a team needs once rules are shared.
- [The Karpathy CLAUDE.md: what is in it and does it help?](https://gethrbr.com/blog/karpathy-claude-md): The Karpathy CLAUDE.md is 65 lines of four coding principles, written by a developer from Karpathy’s January 2026 post. What it says, what it cannot do.
- [AGENTS.md vs CLAUDE.md: which tools read which](https://gethrbr.com/blog/agents-md-vs-claude-md): AGENTS.md is the file Codex, Cursor, Copilot and Antigravity read; CLAUDE.md is Claude Code's. Who reads which in October 2026, and one source of truth.
- [Cursor rules best practices, and why rules stop working](https://gethrbr.com/blog/cursor-rules-best-practices): Cursor rules live in .cursor/rules as .mdc files. Which mode to use, how to share them with a team, and a checklist for rules that do not fire.
- [How to write copilot-instructions.md for GitHub Copilot](https://gethrbr.com/blog/copilot-instructions-md): copilot-instructions.md gives GitHub Copilot repo-wide instructions. How to write it, path-specific .instructions.md files, and which surfaces read AGENTS.md.
- [Why does Claude ignore my CLAUDE.md?](https://gethrbr.com/blog/why-claude-ignores-claude-md): CLAUDE.md is context, not enforcement. Rules lose when they never loaded, conflict, go stale, drown in irrelevant ones, or sit deep in a long session.
- [How long should CLAUDE.md be?](https://gethrbr.com/blog/how-long-should-claude-md-be): Anthropic says under 200 lines. The file is in context for the whole session, so cut what the agent can read from the code and keep commands and gotchas.
- [Should you commit CLAUDE.md and the .claude folder?](https://gethrbr.com/blog/should-you-commit-claude-md): Commit CLAUDE.md, settings.json, rules, skills, agents and .mcp.json. Keep CLAUDE.local.md and settings.local.json out of git. What git still does not solve.
- [Claude Code memory vs Codex, Cursor and Copilot memory](https://gethrbr.com/blog/claude-code-memory-vs-codex-cursor-copilot): Claude Code and Codex keep memory on your machine, Copilot Memory per repository, and Cursor dropped Memories for rules. None holds your team's decisions.
- [CLAUDE.md vs skills vs hooks vs rules vs subagents](https://gethrbr.com/blog/claude-md-vs-skills-vs-hooks): Where each team rule belongs in Claude Code: CLAUDE.md for every session, path rules for one area, skills on demand, hooks for must-never. Plus Cursor.
- [Why does my AI agent keep making the same mistake?](https://gethrbr.com/blog/ai-agent-keeps-making-the-same-mistake): Your correction lived in one chat, and the next session starts fresh. Write corrections down as rules with reasons; use a hook for what must never run.
- [How to turn PR review comments into agent rules](https://gethrbr.com/blog/code-review-comments-as-agent-rules): Keep review comments the author accepted and that recur, rewrite each as a short rule with its reason and PR, and put it where every coding agent loads it.
- [How to get Slack decisions into AI coding agents](https://gethrbr.com/blog/slack-decisions-for-ai-coding-agents): Coding agents never see Slack decisions unless someone pastes them or the agent searches. Capture them once, approve them, and serve them to every agent.
- [What is a company brain?](https://gethrbr.com/blog/what-is-a-company-brain): A company brain captures what a team decided where it decides, gets it approved, keeps it current, and serves the relevant facts to people and AI agents.
- [Context engineering for Claude 5: what changes for teams](https://gethrbr.com/blog/context-engineering-after-claude-5): Anthropic cut 80% of Claude Code's system prompt with no eval loss. For teams: a short CLAUDE.md of gotchas, judgment over rules, skills, and a gap in memory.
- [Context rot and stale rules in CLAUDE.md](https://gethrbr.com/blog/context-rot-stale-rules): Context rot is accuracy falling as input grows. Stale rules, like renamed paths and reversed decisions, are worse: relevant, wrong, and sent every turn.
- [Is AGENTS.md useful? What the research found](https://gethrbr.com/blog/is-agents-md-useful): Hand-written AGENTS.md files help a little, generated ones hurt a little, both cost about 20% more. What ETH, McMillan, Vercel and rule studies found.
- [Architecture decision records (ADRs) for AI coding agents](https://gethrbr.com/blog/architecture-decision-records-for-ai-agents): Agents follow an ADR only if it is in context when they write the code. Keep ADRs short in the repo, index them in AGENTS.md, and mark superseded ones.
- [What is AI agent memory poisoning?](https://gethrbr.com/blog/agent-memory-poisoning): Untrusted content written into an agent's persistent memory that steers later sessions. OWASP lists it as ASI06. Stale memory is the daily form.
- [The best MCP memory servers for teams (2026)](https://gethrbr.com/blog/mcp-memory-server-for-teams): Harbor, Supermemory, Basic Memory, Zep, Mem0, Cognee and Hjarni, checked against their own docs: who writes, what is reviewed, and what each costs.
- [How to keep the same rules in Claude Code, Codex and Cursor](https://gethrbr.com/blog/share-memory-claude-code-cursor-codex): One AGENTS.md covers a repo. Across repos, each tool keeps its own global rules, and none reads the others. The options for you, and for a team.
- [Why a growing CLAUDE.md stops working](https://gethrbr.com/blog/why-growing-claude-md-stops-working): A CLAUDE.md is in the prompt on every call. ETH Zurich found a developer-written context file moved success 2.4 points, not significant, at 20% more cost.
- [Which of my rules did Claude Code actually use?](https://gethrbr.com/blog/served-vs-cited): Claude Code shows which CLAUDE.md and rules files loaded, not which it used. What /context and the InstructionsLoaded hook show, and what served vs cited adds.

## Comparisons

- [Harbor vs Mem0](https://gethrbr.com/compare/mem0): Mem0 remembers users for the agents and apps you build. Harbor serves your team's decisions to Claude Code, Cursor, Codex and Copilot. Side by side.
- [Harbor vs Zep](https://gethrbr.com/compare/zep): Zep builds governed context graphs from business data. Harbor serves an engineering team's decisions to its coding agents. Side by side.
- [Harbor vs Unblocked](https://gethrbr.com/compare/unblocked): Unblocked answers on demand from your code, Slack and docs. Harbor serves your team's decisions to every coding session from the first turn. Side by side.
- [Harbor vs ByteRover](https://gethrbr.com/compare/byterover): ByteRover is shared memory your agents curate as files. Harbor learns your team's decisions from Slack, PRs and docs and serves them. Side by side.
- [Harbor vs Cognee](https://gethrbr.com/compare/cognee): Cognee is an open-source memory engine with a Company Brain. Harbor is a hosted brain with a gate you set and served and cited counts. Side by side.

## Pages

- [Home](https://gethrbr.com/): what it is, what your agents get, who decides what gets in, and what it counts.
- [Docs](https://gethrbr.com/docs): the playbook, and which of its pages is yours.
- [Rule-file checker](https://gethrbr.com/check): paste a CLAUDE.md, AGENTS.md, Cursor rule or Copilot instructions file and see which agents load it and which of its lines no agent enforces. Runs in the browser.
- [Pricing](https://gethrbr.com/pricing): the three plans, what is never metered, and the FAQ.
- [Contact](https://gethrbr.com/contact): reaches the people building it, not a sales team.
- [Terms of Service](https://gethrbr.com/legal/terms)
- [Privacy Policy](https://gethrbr.com/legal/privacy)
- [Data Processing Agreement](https://gethrbr.com/legal/dpa)
- [Refund Policy](https://gethrbr.com/legal/refund)

## Contact

hello@gethrbr.com