Orrery.

Architecture walkthrough

How the loop works

Orrery is not a repo to clone — it is a hosted loop between discovery and execution. Follow the four steps below with live consoles and API endpoints; each section links to proof you can run today.

1 · Gaze

Browse the shelf — no tool payloads

Gaze returns a bounded shortlist: names, facets, oracle pills, and prices. The agent is the semantic router — the shelf never ships full tool schemas or execution payloads.

2 · Resolve

Skill DNS — name to endpoint and digest

Resolve turns a skill name into the MCP URL, content digest, and public key id. This is the binding step before call — not a Git checkout.

3 · Call

Publisher-direct — same contract everywhere

Call the publisher you resolved — Cursor, Claude Code, CI, or a browser console. Direct star MCP endpoints carry the teaching trio; default /mcp is slim discovery plus one call_skill forwarder. Publisher mounts stay canonical.

Teaching trio — point, don't clone. Offline copies cannot mint fresh UTC or re-observe upstream.

4 · Seal

Signed Envelope — verify before trust

Every bounded call returns an Ed25519-signed Chirp Envelope. Verify the signature against published keys before you treat the payload as evidence. Constellation runs chain step receipts into a composite envelope.