COMPASS documentation

The operating layer for companies where people and AI agents work together. Agent memory that knows what still holds.

People lead. Agents work in every department, by your rules: each agent holds a position with a brief and limits, proposes rules that a person confirms with one click, hands work to others as tasks, and shares one memory in which every fact and rule says since when it holds and why.

This documentation is for anyone connecting an agent (Claude Code, Claude Desktop, claude.ai, Cursor, Codex, or a custom MCP client) or a service to COMPASS, and for developers working directly with the REST API. Quickstart, examples and the Claude Code plugin are on GitHub: github.com/Athron-new/compass-agents.

Where to start

• Quickstart — start with the sample company Nordlicht and connect Claude in five minutes. Then the developer path in five steps: import a rule set, ask what holds and why, look at deviations, supersede a rule and watch the old one stay visible as superseded.
• Concepts — the one form COMPASS stores everything in: Sache, Aussage, bi-temporal validity, superseding, the one-current-rule guarantee, tenants, and personas.
• Guides — task-focused walkthroughs: importing a rule set, setting and superseding a rule, attaching a check, connecting an editor, searching.
• Reference — the REST API grouped by area, every MCP tool with its parameters, and the shared form each record is built from.

Scope of this documentation

Everything documented here is checked against the running code. An endpoint or tool is described only if it exists both in our specification and in the running service; everything else is left out.