How to Add a Custom MCP Server to Claude and ChatGPT

Connecting a custom MCP server to Claude or ChatGPT’s standard chat interface is possible in both, but the two platforms make wildly different assumptions about how much friction that connection should require — and the gap tells you something about how each company thinks about agent trust.

The Framework

Both platforms expose the same underlying idea: a chat interface that can call out to an external MCP server for tools or data beyond what’s built in. But “add a connector” means something different depending on which chat app you’re in — one treats it as a first-class, always-visible feature; the other buries it behind a security warning and a plugin-directory workflow inherited from an earlier product era. That gap isn’t an oversight on either side. It’s a visible artifact of two different product histories: Claude’s Connectors menu was built for MCP from the start, while ChatGPT’s plugin infrastructure predates MCP as a standard and is being repurposed to carry it, warnings and all.

Claude: Toggle It On From the Connectors Menu

In Claude, adding a custom MCP server takes four steps: open the + menu next to the chat prompt, go to Connectors → Add custom connector, fill in a name and the MCP server’s URL, and optionally configure authentication under Advanced settings. Once added, the connector shows up in the Connectors menu with a simple on/off toggle — you can disable it for a conversation without removing it entirely, which matters if you’re testing a server you don’t fully trust yet.

ChatGPT: Developer Mode, a Plugin Directory, and a Risk Warning

ChatGPT’s path runs through infrastructure built for an earlier plugin system. You first have to enable Developer mode in Security and login settings — a toggle explicitly marked “ELEVATED RISK.” From there, you go to the Plugins directory, click the + icon, and fill out a “New Plugin” modal with a name, description, and server URL, setting Authentication to “No Auth” if the server doesn’t require it, then acknowledging the risk warning to finish creating it. To actually use it in a conversation, you click the + icon next to the chat input, search for the MCP by name, and accept a connection prompt the first time it’s invoked — a second discovery step Claude skips entirely.

Where It Doesn’t Apply

Neither platform’s documentation makes this obvious, but MCP servers set up this way aren’t available everywhere in the app. They don’t show up as usable tools inside ChatGPT’s Advanced Voice Mode, so a workflow wired up for text chat can silently stop working the moment a conversation shifts to voice — something worth testing explicitly rather than assuming will carry over, following the same pattern Stack Archive has flagged in AWS’s own MCP server rollout: the protocol being standardized doesn’t mean every surface implements it identically yet.

If you’re standardizing MCP access across a team that uses both tools, budget for that gap explicitly: a setup guide that works for Claude users will not transfer cleanly to ChatGPT users without separately walking them through Developer mode and the risk acknowledgment, and skipping that step is the most common reason a “just add the server” instruction stalls out in practice.

Source: Simon Willison (TIL)

Content created with AI assistance and reviewed for accuracy.