ChatGPT

Setup guide only, not run live

What Zyx serves and which gates apply: Use Zyx from your AI tool.

ChatGPT runs in OpenAI’s cloud. It cannot reach a Zyx on a private network and cannot send a key header, so it needs the public OAuth path.

Before you start: the operator publishes the MCP paths

The operator makes four paths public: /api/mcp, /.well-known/oauth-protected-resource, /.well-known/oauth-authorization-server and /oauth. Nothing else of Zyx is made public.

Setup

1. In ChatGPT’s settings, turn on developer mode for connectors and create a custom connector. The menu names change; OpenAI’s help pages have the current path. A Business or Enterprise workspace may need its admin to allow it.
2. URL https://zyx.example.com/api/mcp, authentication OAuth. ChatGPT registers itself with Zyx and opens Zyx’s consent screen.
3. Sign in to Zyx, tick the scopes (read, and tools if ChatGPT may change things) and allow. The grant shows on the keys screen labelled OAuth; revoke it there. Zyx’s change log records every grant, revoke and replayed token.
4. ChatGPT has no field for a folder. The owner’s connector works in the company root. A staff member’s runs in their own private session: their own folders and sign-ins answer there, and every approval it asks for is theirs alone.

What ChatGPT calls

ChatGPT’s connectors and deep research call two read-only tools by name, search and fetch. The read scope gives both. search returns the files and synced items this person may read, each with an id, a title, a link into Zyx and the matching passage. fetch reads one whole item by its id. Both use the same readers as Zyx’s own search, so the person’s folders, the items’ readers and the privacy gate all apply.

What was proven

• Not proven with ChatGPT: no ChatGPT connector has been created against Zyx.
• The server side (discovery, registration, PKCE consent, tokens, a tool call as that person) is proven by an automated test with the MCP SDK’s own OAuth client. search and fetch have their own tests.