ChatGPT MCP client

Does your MCP server work in ChatGPT?

OpenAI's assistant, including ChatGPT apps. ChatGPT renders MCP apps, and handles tool results its own way. Our catalog records the details below. Test your server against it with MCPJam.

Test against ChatGPT View ChatGPT capabilities

Terminal

npx @mcpjam/inspector@latest

What ChatGPT supports

MCP apps

ChatGPT renders MCP apps, placing tool-result media inline.

Tool-result images

The model sees images returned via direct content, embedded resources, and linked resources. That covers every result shape we track.

Display

Runs inline by default; supports inline and fullscreen and pip on desktop.

MCP extensions

Advertises io.modelcontextprotocol/ui.

Emulation model

MCPJam emulates ChatGPT using openai/gpt-5-nano, so tests reflect how this host actually drives your tools.

How we know

Vendor docs: Based on the vendor's published documentation. Last reviewed 2026-07-08.

Test your server against ChatGPT

Test an MCP app in ChatGPT

Walkthrough for test an MCP app's UI in ChatGPT.

Test MCP tools in ChatGPT

Walkthrough for test MCP tools and their results in ChatGPT.

Connect a server to ChatGPT

Walkthrough for connect an MCP server to ChatGPT and verify it.

ChatGPT capability questions

Does ChatGPT support sampling?

Live support matrix for MCP sampling, per client.

Does ChatGPT support elicitation?

Live support matrix for MCP elicitation, per client.

Does ChatGPT support roots?

Live support matrix for MCP roots, per client.

Compare MCP host capabilities side by side

The full Can I Use matrix on caniuse.dev.

How to test your MCP server against ChatGPT

Run npx @mcpjam/inspector@latest, connect the server config you already use, and select ChatGPT. MCPJam drives your server the way this host does, including its model (openai/gpt-5-nano), so you see the tool calls it actually makes instead of guessing from a bug report.

From there, Trace shows every step from prompt to response, the OAuth Debugger shows where authorization breaks, and Evals score whether the model picks the right tool across realistic requests. All of it runs locally, free and open source.

Frequently asked questions

Yes. Our catalog records ChatGPT as rendering MCP apps, with tool-result media placed inline. (Vendor docs, last reviewed 2026-07-08.)

The model sees images returned via direct content, embedded resources, and linked resources. That covers every result shape we track.

Run npx @mcpjam/inspector@latest, connect your server, and select ChatGPT. MCPJam emulates the host with openai/gpt-5-nano so you can watch the real tool calls, trace them, and score them with evals.

It comes from MCPJam's host-compat catalog (the same data behind caniuse.dev), recorded as "vendor docs" and last reviewed 2026-07-08. Based on the vendor's published documentation.