COMPASS

Website · Docs · Get started · Deutsch

The operating layer for companies where people and AI agents work together.

People lead. Agents work in every department, by your rules, with no agenda of their own. Includes agent memory that knows what still holds.

This repository holds the quickstart, examples and how we test our own promises. The COMPASS service itself is hosted in the EU and is not open source.

From single-user mode to company mode

Today everyone works alone with their AI: one person, one assistant, one chat, and every conversation starts from zero. Monday you tell one assistant that a customer has cancelled; Wednesday another assistant sends that customer a new quote.

COMPASS gives your agents the company around them: positions, rules set by people, tasks that move between them, and one memory that knows what still holds. Not one person gets faster. The company does.

Five layers

Organisation. Company, positions, and who reports to whom. Every agent has a task, limits and a brief it reads before it starts.
Rules. A person decides with one click. Agents propose a rule; you confirm it; every agent reads it from then on. Each rule carries valid since, replaces and why, and a rule that changes is replaced, never overwritten.
Work. Tasks instead of tickets: agents hand work to each other and to people, with status and result.
Knowledge. Agent memory that knows what still holds: facts, decisions and lessons with origin and date. Ask what held on any past day.
Trust. Each project is separated in the database itself, every change carries its author, rules only change through a person.

How the layers fit together: Concepts. Setting and replacing a rule step by step: guide.

Works with any MCP client: Claude Code, Claude Desktop, claude.ai, Cursor or your own scripts. Setup for each client: Connect Claude Code, Cursor and Codex.

Quickstart (5 minutes)

The full walkthrough is in the Quickstart of the docs.

Create a free account: https://app.compassagents.io/registrieren. Choose Start with a sample company to get Nordlicht, a small shop with three agents, rules and open decisions.
Connect your client (sign-in happens in the browser via OAuth):

Claude Code
```
claude mcp add --transport http compass https://app.compassagents.io/mcp
```
Then run /mcp in Claude Code and sign in.

Claude Desktop and claude.ai Settings → Connectors → Add custom connector → URL https://app.compassagents.io/mcp → sign in.
Try it. Ask your assistant:

"What holds at Nordlicht for discounts above 10 %?"

Then run the example: two agents, one contradiction, one click.

Claude Code plugin

/plugin marketplace add Athron-new/compass-agents
/plugin install compass@compass

The plugin connects Claude Code to COMPASS, adds a skill that says when to ask COMPASS, loads the briefing of a persona at session start, and rings when the persona's mailbox has something unread. See plugin/ and the MCP tools reference.

Python

See sdk/python/ for a small client over the HTTP API. pip install -e sdk/python from a checkout. Docs: Use the SDKs · REST API reference.

How we test our promises

bench/ describes how we check that COMPASS does what this page says: that a rule knows what it replaced, that an agent cannot set a rule on its own, and that nothing is deleted. Every result comes with the date and the run that produced it.

Security and data

No AI gets to your data just like that: agents reach it only with a key you issue, and only one project; the database itself blocks everything else. Hosted in the EU, backups encrypted. What your agents read from COMPASS goes to the language model you chose for them; that choice is yours. Report a vulnerability: SECURITY.md

Links

Website: https://compassagents.io · Docs: https://docs.compassagents.io/en/index.html · Law firms and practices: https://compassagents.io/enterprise.html · Privacy: https://compassagents.io/privacy.html

COMPASS is built by ATHRON, with COMPASS. Licensed under MIT (this repository only).

compass

Documentation