Contact us
(Guide)

How to test an MCP serverin ChatGPT and Claude

Updated · 4 min read

A server that works locally doesn't always work in ChatGPT and Claude. Both connect to it from their own cloud, sign the user in through OAuth and pick the tools themselves. Here is the testing order we use for every server, and the errors that come up in it most often.

Testing order

  • MCP Inspector on your machine: protocol, tools, sign-in.
  • A public HTTPS URL through a tunnel, because ChatGPT and claude.ai can't connect to localhost.
  • Claude and ChatGPT: whether the model picks the right tools and understands the answers.
  • Before you submit to the directories: a test account, every tool in a fresh conversation, error messages.

MCP Inspector

The Inspector is the official tool for testing MCP servers. Start it with "npx @modelcontextprotocol/inspector" and open the URL it prints, token included, in your browser. For a remote server, pass "--server-url https://your-domain/mcp --transport http".

  • Tools, resources and prompts: call them by hand and see exactly what the model gets, annotations included.
  • Sign-in: when the server answers 401, the Inspector runs the OAuth flow itself, with client registration through DCR or CIMD.
  • UI: the Apps tab renders MCP Apps views in a sandbox.
  • Raw JSON-RPC messages, HTTP headers and server logs.
  • A "--cli" mode for scripts and CI, for example to check the tool list after every deploy.

A public URL

ChatGPT and claude.ai connect to your server from their cloud, so you need a public HTTPS URL. To test a local server, a tunnel is enough, for example Cloudflare Tunnel or ngrok. Claude connectors connect over IPv4 only.

If your server checks the Host header (createMcpExpressApp from the TypeScript SDK does), add the tunnel's hostname to the allowed hosts. Otherwise you get a 403 "Invalid Host" error. ChatGPT also offers Secure MCP Tunnel for connecting a server that isn't public.

Claude

  • claude.ai: Customize > Connectors > Add custom connector. The free plan allows one custom connector, Pro and Max more.
  • On Team and Enterprise, an organization owner adds the connector in the organization settings.
  • You can change a connector's name and URL, but not its sign-in settings. To change those, remove the connector and add it again.
  • Claude Code: "claude mcp add --transport http name url", then the /mcp command shows the status and lets you sign in.
  • Limits on claude.ai and the desktop app: a tool result of up to about 150,000 characters and 240 seconds per call. Claude Code accepts up to 25,000 tokens by default (the MAX_MCP_OUTPUT_TOKENS variable).

ChatGPT

  • On chatgpt.com/plugins choose +, then Add custom MCP server. Enter the URL including /mcp, set up sign-in and create the plugin. In a new conversation, type @ and pick it from the list.
  • Custom servers are added on the web. Full access, including tools that change data, is on Business, Enterprise and Edu (on Business only admins and owners). On Pro only read tools work.
  • After every change, deploy the server, click Refresh on the plugin page and test in a new conversation, because an old one remembers the previous tool list.

UI in the conversation

ChatGPT and Claude both support the open MCP Apps standard, so the same view (a product card, a cart, a form) works in both. Test two things separately: the view the user sees and the text result the model reads. The view should run with no console errors and restore its state when the conversation is reopened.

The most common errors

  • A 401 response without a WWW-Authenticate header pointing to the resource metadata (/.well-known/oauth-protected-resource). Without it, sign-in never starts.
  • No PKCE with the S256 method. ChatGPT and Claude both require it.
  • A redirect to another host drops the Authorization header and ends in a 401.
  • A firewall or CDN blocks Anthropic's addresses. You then see 403 errors your app never sent.
  • OAuth endpoints take longer than 10 seconds to answer and the connection times out.
  • Overly strict Origin checks reject Claude's requests.
  • Tools without a title and a readOnlyHint or destructiveHint annotation. Claude's directory won't accept them. The full list of requirements is in our submission checklist.
FAQ
Do I need a paid plan to test in Claude?
No. The free plan lets you add one custom connector. Submitting to the directory does require a paid plan.
The server works in the Inspector but not on claude.ai. Why?
Usually because the Inspector connects from your machine and claude.ai from Anthropic's cloud. Check that the URL is public and reachable over IPv4, that no firewall blocks the traffic and that the server accepts the tunnel's host.
Can testing be automated?
Yes. The Inspector's "--cli" mode calls methods such as tools/list and tools/call and returns JSON, so it fits into CI. How the model picks tools still has to be checked in real conversations.
Guides

Need help with an MCP server?

MCP servers

LET'S MAKESOMETHINGPEOPLE→REMEMBER

Get in touch