Your Shinobi, inside ChatGPT and Gemini
Customers already ask assistants about your business. Publish gives those assistants a real source: one key-authed endpoint that answers only from your live knowledge base, cites the articles it used, and says an honest “I don't know” instead of letting someone else's model guess your policies.
Turn it on, paste two things, done
No SDK, no webhook choreography. Mint a key on the Publish page, import the OpenAPI link into a Custom GPT, paste the key into its auth settings, and the GPT answers with your knowledge from that moment on.
A Custom GPT in minutes
The Custom GPT builder imports our OpenAPI spec from one URL. Your publish key goes in its auth settings, and every product or policy question routes through your agent.
The same schema, as a function
Gemini consumes the endpoint as a function declaration: same URL, same key, same grounded answers. Nothing to rebuild per assistant.
Any API-calling assistant
Claude, agent frameworks, an internal tool, a plain curl: one POST with a Bearer key returns the answer, whether it was answered, and the sources.
What the assistant relays is what your knowledge says
Every call searches your live articles first. The answer is written only from what it finds, in your voice, and the response names the articles it used so you always know where a claim came from.
When the knowledge base has no answer, the response says answered: false with a plain-language line the assistant can relay honestly. Your policies never get invented by someone else's model.
One key, shown once, revocable in one click
Publishing is off until you turn it on. The key is displayed once at minting and Shinobi stores only a fingerprint; rotate it and the old key dies instantly, everywhere.
Prove it before you ship it
The console's test panel runs your question through the real public path: the same key check, the same engine. What you see is exactly what ChatGPT relays.
Every call accounted for
Publish calls land in Usage & Metering under their own surface, with tokens and latency, next to everything else your agent does.
Off means off
Turn publishing off and the endpoint rejects every call immediately. Rotate the key and every assistant using the old one stops cold.
What is live and what is next
Publish ships in the open, and this page updates the day the next piece does, not before.
Live today: the key-authed endpoint, the importable OpenAPI spec, grounded answers with sources, honest no-answers, per-call metering, and the console test panel.
Live today: the same agent as a Model Context Protocol server. Claude and MCP-native assistants add the MCP URL as a connector with the publish key as the Bearer token and get one tool, ask_support_agent, grounded only in the live knowledge base.
Publish answers from the same brain Studio builds and Self-Healing keeps improving. Publish once, and every fix you approve makes every assistant smarter.