TLDRocket
Sign in

Adding a custom MCP server to Claude and ChatGPT

Simon Willison's Weblog Simon Willison

Simon Willison figured out how to hook a custom MCP server into the regular Claude and ChatGPT web chat interfaces.

Based on reporting by Simon Willison's Weblog, Simon Willison — read the original for the full story.

Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error

Simon Willison has been digging into the Model Context Protocol lately, and his latest find is a practical one: both Claude's and ChatGPT's standard web chat interfaces can actually talk to custom MCP servers. The catch, as he lays out in his post from 29th July, is that neither company makes this obvious. There's no big "add your MCP server here" button waiting for you. You have to go looking.

Willison's writeup is one of his classic TIL (Today I Learned) posts, the kind where he works through a fiddly setup process step by step so the rest of us don't have to. The gist is that connecting your own MCP server to these mainstream chat UIs is entirely doable, but it takes a fair number of steps to get there, none of which are particularly well signposted by Anthropic or OpenAI. This matters because MCP has quickly become the de facto way for LLM tools to reach out and grab context or capabilities from external systems, and the web chat interfaces are where most people actually spend their time, not in some developer sandbox.

What's notable here is less the specific instructions and more the pattern Willison is pointing at: MCP support is arriving in these products before the onboarding experience has caught up. Anthropic and OpenAI have clearly wired the plumbing in, but the switches and knobs are buried, undocumented in any user-friendly way, or just assumed to be something only power users will bother with. That's a familiar shape for anyone who's watched AI features roll out over the past couple of years — capability first, usability later, if at all.

This is also part of a broader thread in Willison's recent writing, where he's been circling MCP from multiple angles, including how models handle tool use and what happens when that access goes wrong, as in his recent piece on OpenAI's mishap with Hugging Face. Getting a custom server connected to the chat UI is a small, almost mundane task by comparison, but it's the kind of connective tissue that determines whether MCP becomes something ordinary users benefit from or stays a protocol mostly discussed by developers on forums.

My take — AI-written commentary, not fact-checked reporting

I'll say it plainly: if a documentation-obsessive like Willison has to write a TIL post just to wire up a basic MCP connection, the UX here is broken, full stop. Protocols win when the plumbing disappears into the background, not when explaining the plumbing becomes content. Anthropic and OpenAI shipped the wiring and skipped the doorbell.

Read more about this at: Simon Willison's Weblog

Related stories

The daily briefing

Every AI story that matters, in your inbox by 8am.

TLDRocket reads all relevant sources, removes duplicate coverage, and summarises the day in two minutes. Follow companies and topics for alerts, or get the briefing in Slack. Free, no spam, unsubscribe anytime.