Built an MCP integration for ChatGPT in 1 week - here's what the docs don't tell you

Then OpenAI adopted MCP, and we wondered what would happen if we went where the questions already are.

We just submitted mindPick to ChatGPT's app marketplace. Here's what we learned, including the things we had to figure out ourselves.

What is MCP?

Model Context Protocol. A standard that lets any app connect with AI assistants. Anthropic created it, and OpenAI adopted it.

You develop one server that uses MCP. Now ChatGPT, Claude, Cursor, and eventually others can call your app when it's relevant.

Why this matters for indie hackers:

ChatGPT has over 200 million weekly users. You can't buy that reach. You can't SEO your way there.

But you can integrate for free.

The barrier isn't money. It's knowing this exists and building the connector.

What we built:

- MCP server: a small backend explaining what our app can do
- Three tools ChatGPT can call: search, fetch, start_question
- Widget: displays expert cards inside the chat

Build time: about one to two weeks. Stack: React, Vercel, Xano. Cost: $0.

What the docs don't tell you:

1. You need both search and fetch tools. If you miss either one, you get a "search action not found" error. This isn't clear from the MCP spec; it's a ChatGPT-specific requirement.

2. Developer Mode is hidden. Go to Settings, then Connectors, then Advanced, and finally Developer Mode. It took us 20 minutes to find it.

3. Docs are scattered. The MCP spec lives on Anthropic's site. ChatGPT-specific requirements are on OpenAI's site. You need information from both.

4. Tool descriptions serve as your "SEO." ChatGPT decides when to call you based on matching your tool descriptions to user intent. There is no documentation on the algorithm. We treat it like writing meta descriptions, focusing on specific problems rather than generic features.

5. Security is crucial. OpenAI warns that "custom connectors are not verified. Malicious developers may attempt to steal your data." If you work with real data, treat this project's security seriously.

The catch we're still figuring out:

Being integrated doesn't mean being surfaced. There is a cold-start problem. If ChatGPT calls us and we return no relevant expert, that's a poor experience. It might learn not to call us again.

Where we are now:

- ChatGPT marketplace: in review
- Cursor marketplace: live (proving the MCP spec is portable)
- Claude Desktop: ready

We have no idea if this drives meaningful traffic yet. We will report back with data.

If you want to try this:

- MCP spec docs
- OpenAI's integration guide
- Our integration page: mindpick.me/integrations

Questions:

Has anyone else built MCP integrations? What made ChatGPT actually surface your app, and what didn't work?

And for those who tried and gave up: what was the blocker?