One MCP server, four assistants: what broke on Claude, ChatGPT, Gemini and Grok
quiet_flare · reddit · 2026-11-14
The author tested one remote MCP server (OAuth, custom connector) across Claude, ChatGPT, Gemini and Grok and found the protocol identical but behavior very different:
- Schema adherence varies wildly: typed list blocks came back with wrong types, answers as letters or raw text, missing ids, or whole lists as one JSON string. Fix: a repair step before validation — alias tables, string parsing, numbering missing ids, with a strict "never invent content" rule.
- Write error messages for the model, not developers: name the block and field, show one correct example, and end with "correct this and call again now, with the same content" — enabling one-shot self-correction.
- Server instructions load into every conversation: an early version that told assistants to call a tool up front caused irrelevant chats to trigger the server.
- Put rules in three places (instructions, prompt, resource) since clients read different ones.
- Polish details: title every tool, redirect from the OAuth grant page, and expect new connectors to only appear in fresh chats.
Also, connector setup is mobile-only on Claude and Grok, web-only on ChatGPT and Gemini.
More from coding & agent
- Solo dev builds real-time conference translation platform for 70 languages with an agentic harness — evisoft · 2026-11-14
- Open-source Blender add-on runs NVIDIA Kimodo, ARDY and MotionBricks locally for AI motion generation — bdsqlsz · 2026-11-14
- TestPrune: pruning regression tests to unblock AI-driven autonomous bug fixing — burkov · 2026-11-14
- Context beats model choice: Databricks benchmark shows 91% vs 58% for data agents — CautiousUse8597 · 2026-11-14
- Developer skips localhost entirely: prompts go straight to production with AI agent safeguards — marclou · 2026-11-14
- HeyGen MCP demo: one ChatGPT prompt books tickets and sends an avatar birthday video — toolstelegraph · 2026-11-14