Simon Willison recently walked through connecting a custom MCP server to both Claude and ChatGPT's standard chat interfaces, and the takeaway is more instructive than the specific steps: it works, but it isn't yet the kind of thing a non-technical user could do on their own.
This article covers what the Model Context Protocol (MCP) actually is, why extending a mainstream chat interface with a custom data source still requires real configuration work, and what that friction implies about where these products are headed.
What MCP actually connects
MCP is a protocol — a standardized way for AI assistants to talk to external tools and data sources — that lets a model call out to something beyond its own training data or the chat app's built-in features. Think of it like a universal plug adapter for AI assistants: instead of every tool needing a bespoke integration built by the AI vendor, a tool that speaks MCP can be plugged into any assistant that also speaks MCP.
That standardization is the whole point. Before protocols like this existed, adding a new capability to a chat assistant meant either waiting for the vendor to build first-party support or building a narrow, assistant-specific integration that wouldn't transfer to a competitor's product. MCP is explicitly designed to break that lock-in — a server built once should, in principle, work with Claude, ChatGPT, or any other MCP-compatible client.
Why "technically possible" still means "several steps"
Willison's write-up is candid that hooking a custom MCP server into Claude's or ChatGPT's standard consumer chat interface — not a developer SDK, the actual chat app most people use — requires multiple configuration steps. That's a meaningful gap between what the protocol enables in principle and what a typical user can do in practice today.
A protocol being open and well-specified doesn't automatically make it accessible — the last mile from "technically supported" to "a normal user can set this up" is often the hardest part.
This is a familiar pattern in developer tooling generally: the protocol layer matures faster than the onboarding experience built on top of it. Early browser extensions, early webhooks, early OAuth flows all went through a similar phase where the underlying mechanism worked but the setup process assumed enough technical comfort that it effectively gatekept the feature to developers.
What this signals about where these products are headed
The friction here isn't a permanent property of MCP — it's a snapshot of where consumer AI chat interfaces are in their extensibility maturity. Both Claude and ChatGPT clearly see external tool connectivity as strategically important; the fact that it's supported at all, even clumsily, in the mainstream chat product (not just a developer-only surface) suggests the vendors are racing to make this simpler, not treating it as a niche feature for power users.
The practical prediction that follows: the configuration steps Willison documents today are likely to shrink over the next several product cycles, the same way installing a browser extension went from editing manifest files to a one-click store install. The protocol is stable; the UX around it isn't finished yet.
What this means for builders
If you're building a tool meant to plug into AI chat assistants, MCP is the standard worth building against rather than a proprietary per-vendor integration — the setup friction is a UX problem the platforms are actively working to solve, not a sign the protocol itself is the wrong bet. Building early against MCP means your tool is positioned to benefit as onboarding improves, rather than needing to be rebuilt once it does.
For teams evaluating whether to give end users access to custom tools inside Claude or ChatGPT today, budget for real onboarding documentation and support — this isn't yet a drop-in feature you can hand to a non-technical user with a single link.
Conclusion
MCP working end-to-end across Claude and ChatGPT is a genuine milestone for interoperability between competing AI platforms. But the multi-step setup Willison documents is a reminder that a protocol being open and functional is a different milestone from a protocol being easy — and for AI chat assistants aiming at mainstream adoption, that second milestone is the one that will actually matter to most users.