Skip to content

openarity

Capability is not a flat list.

An open-source agent platform that selects capability by walking a graph rather than ranking a list — and an enterprise harness around the run.

A swarm of agents joined by edges converges into the openarity mark, a rounded-square O; it splits on the diagonal into two halves, a visor opens, a signal runs down the cut from one node to the other, and the name openarity arrives beside it A swarm of agents joined by edges converges into the openarity mark, a rounded-square O; it splits on the diagonal into two halves, a visor opens, a signal runs down the cut from one node to the other, and the name openarity arrives beside it

Get started

Run the whole stack with one command, put a brain in front of a Postgres you already have, or take the agent loop on its own. Read the docs.

  • Traversal, not ranking: Tool selection measurably collapses as the pool grows, and retrieving tools by embedding only postpones it — the paper that popularised that fix reports its own precision degrading into the thousands. What separates two similar tools is not in their descriptions. It is what they connect to.
  • Graph engineering, not GraphRAG: A knowledge graph models what a system knows. This one models what it is — agents, tools, skills, teams and learnings, with edges of authority and provenance. Asking what a run may do and what it can do is the same walk.
  • Boundaries that fail loudly: The agent loop cannot authorise anything: it is a separate module, so the import does not compile. It cannot reach a provider either — a CI check fails on the linked package. Rules a reviewer has to remember last until the first deadline.
  • Deletion that actually deletes: Attachments are encrypted with a per-team key. Deleting a team destroys the key first, so every one of its attachments is unreadable before a single object has been removed — because no transaction spans Postgres, an object store and a vault.
  • One graph, many surfaces: Code, Work, Knowledge, Design, Mobile. Not products that share a vendor — views onto the same graph, so a learning earned in one is capability in every other, with nothing to integrate. None of it is built.
  • An agent loop you can take on its own: sdk/agent is a separate Go module with one dependency and no database, server or configuration file. ReAct, plan-then-act, ReWOO and reflection, behind an interface you can add your own pattern to.

Early

There is no release yet, and no stable API. Every page says what is built and what is not.