OpenAI's MCP Review Rules: Every Tool Needs Annotations, but Hints Aren't Security

shoumikgoswami · reddit · 2026-10-31

InfoQ reported OpenAI engineer Thomas Sarlandie's talk at AGNTCon/MCPCon: ChatGPT's plugin directory rejects MCP servers unless every tool carries annotations, enabling fast-track of read-only calls and confirmation of destructive ones.

Key criteria:

A subtle rule: justification text in submissions can't override annotations. If the server advertises readOnlyHint: false, calling it "functionally read-only" in the form changes nothing — fix the annotation, deploy, rescan, verify.

The other half: the MCP spec requires clients to treat annotations as untrusted, and OpenAI's docs stress they don't replace server-side authorization, input validation, or confirmation for consequential actions. Accurate hints help clients decide when to ask — they enforce nothing.

Original post →

More from coding & agent

coding & agent channel →