Built an MCP server that connects ChatGPT to async expert Q&A. When someone asks a professional question needing human judgment, it surfaces relevant experts with pricing and response times.

Working in dev mode, submitted to ChatGPT marketplace.

What we learned building it:

1. ChatGPT requires `search` + `fetch` tools - could not find in MCP spec, undocumented until you hit the error

2. Tool descriptions = discovery. ChatGPT matches user intent to your descriptions. No algorithm docs. Treat it like SEO.

3. Dev Mode is buried: Settings → Connectors → Advanced → Developer Mode

4. Docs are split: Anthropic owns MCP spec, OpenAI owns ChatGPT-specific requirements. You need both.

5. Cold start problem applies: if you return no results, ChatGPT learns not to call you

Stack: TypeScript, Vercel, Xano

Happy to share more on the build if useful.