ChatGPT connectors and apps

Support

Protocol revision

Not published, and there's no client source to check. The signals conflict: OpenAI's authorization and tool-annotation pages link 2025-11-25, while its lifecycle and tool-result pages link 2025-06-18.

Caveats and sources

Each one names the column it qualifies and cites the page it comes from, which is often not the page in the row's primary source.

Custom headers

Stated outright, and it's the most quotable limit in this matrix: ChatGPT doesn't support machine-to-machine OAuth grants such as client credentials, service accounts, or JWT bearer assertions, and it can't present custom API keys or customer-provided mTLS certificates. It does present an OpenAI-managed client certificate of its own, with published CA chains, so mTLS works in one direction only — you can verify ChatGPT, but ChatGPT can't present your certificate. Source →

Metadata document

ChatGPT prefers a metadata document when one is available, and the builder can choose dynamic registration when both are offered. Its client_id is a URL on chatgpt.com. For token-endpoint authentication it supports none and private_key_jwt, preferring the signed assertion where available. Dynamic registration runs once per server connection, and the registered client is reused. Source →

Sends resource (RFC 8707)

ChatGPT appends resource to both the authorization request and the token request. Its redirect URI is on chatgpt.com, and OAuth configured without offline_access can lose access once the original authorization expires. Source →

stdio

Not directly. ChatGPT connects to remote servers, and a local stdio or HTTP server reaches it only through the Secure MCP Tunnel, which is gated to enterprise customers. Source →

Legacy HTTP+SSE

OpenAI states the requirement positively — a production endpoint must support Streamable HTTP — without saying SSE is rejected, and one sample on its API docs still shows an SSE transport. Unknown rather than absent until one of those pages says so. Source →

Tool metadata

Full MCP support including write actions is in beta and plan-gated to Business, Enterprise, and Edu, with Pro limited to read and fetch. Tool metadata is also frozen at approval: later changes by the developer don't apply until an admin reviews and publishes an update, and a live server that no longer matches the frozen snapshot can error on tool calls. Source →