Multi-agent systems today are bespoke. Every integration is a custom adapter, every handshake is hand-rolled, every capability map is bolted on. The right primitive is the one routing already has — a discoverable, named endpoint, with a capability advertisement, and a standard wire format. zhub takes that primitive and applies it to AI assistants.
Designed an OpenAI-compatible HTTP surface as the wire format — any client that speaks OpenAI's API speaks zhub.
Built service discovery on top of standard mDNS / static registry so an AI's endpoint becomes addressable as a name, not a URL.
Made the skill itself bidirectional — the AI publishing the endpoint can also see who's connected and what they can do, so it can route or delegate.
Kept the install path to three commands. Anything longer dies on the second page of the README.
Published as an open skill — any AI that runs the skills system can pick it up.
Published the protocol and tooling as a public experiment in connecting AI clients through discoverable endpoints.
Explored client-to-agent communication without treating each connection as a separate, hand-built integration.