<MCP.Playground/>
online

Test Your MCP Server With ChatGPT and GPT-5

Free13 GPT models

Paste a server URL and watch GPT-5 call your tools in a real conversation — before you wire it into Developer Mode, the Responses API, the Agents SDK or Codex CLI, each of which fails differently.

Paste a server URL, pick a model, and watch it call your tools in a real conversation. No install.

No server of your own? Leave it blank and use one of the public mock servers.

Why test with GPT?

OpenAI has the broadest MCP surface of any vendor and the most ways to get it subtly wrong. ChatGPT Developer Mode, the Responses API’s mcp tool, the Agents SDK and Codex CLI all reach MCP servers, with different permission models — most notably an asymmetry between read-only and write tools that is easy to miss until a write silently does not happen. Testing the tool calls in isolation first tells you whether a failure is your server, your schema, or the integration path you picked.

How GPT reaches MCP tools

  • ChatGPT Developer Mode and custom connectors, for using a server inside the ChatGPT app
  • The Responses API mcp tool, where you pass your server URL to OpenAI directly
  • The OpenAI Agents SDK, for building your own agent around the server
  • Codex CLI, for local development and scripted runs

OpenAI converts your MCP tools into function definitions, and then each integration path applies its own permission layer on top. The conversion is the same everywhere; the permissions are not. That is why the failure people report as 'GPT ignored my tool' is usually a path-level restriction rather than anything the model decided.

What goes wrong with GPT specifically

  • Read-only and write tools are not treated the same across OpenAI’s integration paths. A server that works fully in the Agents SDK can find its write tools quietly unavailable in the ChatGPT app — check this before assuming your tool is broken.
  • Connector setup in the ChatGPT app is the single biggest source of "my MCP server does not work with ChatGPT" reports, and the cause is usually auth or transport, not the tool definitions.
  • GPT selects tools noticeably differently from Claude on identical schemas — if you developed against Claude, do not assume the descriptions transfer.

Call a write-capable tool on the complex-schema mock here, where no permission layer applies. If it succeeds here and silently no-ops in the ChatGPT app, you have found the read-only asymmetry rather than a bug. All the mock servers →

GPT models available here

ModelModel IDCredits / run
GPT-5.4 nanoopenai/gpt-5.4-nano2
GPT-5.4 miniopenai/gpt-5.4-mini4
GPT-5 miniopenai/gpt-5-mini4
GPT-5.3 Codexopenai/gpt-5.3-codex12
GPT-5.4openai/gpt-5.412
GPT-5openai/gpt-5-chat12
GPT-5.5openai/gpt-5.515
GPT-5.6 Lunaopenai/gpt-5.6-luna6
GPT-5.6 Luna Proopenai/gpt-5.6-luna-pro6
GPT-5.6 Terraopenai/gpt-5.6-terra12
GPT-5.6 Terra Proopenai/gpt-5.6-terra-pro12
GPT-5.6 Solopenai/gpt-5.6-sol15
GPT-5.6 Sol Proopenai/gpt-5.6-sol-pro15

Go deeper

Test with another model

Frequently asked questions

Does ChatGPT support MCP servers?
Yes — through Developer Mode and custom connectors in the app, and through the Responses API mcp tool, the Agents SDK and Codex CLI on the developer side. Each path has its own permission model, which is why a server can work in one and not another.
Why do my write tools not work in ChatGPT?
This is the read-only versus write asymmetry — the most common surprise with OpenAI’s MCP support. Some paths restrict write-capable tools more tightly than read-only ones. Confirm the tool works at all by running it here first; if it succeeds, the problem is the integration path’s permissions rather than your server.
Do I need an OpenAI API key to test here?
No. Runs go through MCP Playground, so you can see exactly how GPT-5 calls your tools without an OpenAI account or key. You need one only when wiring it into your own product.
My server works with Claude but not GPT — what changed?
Nothing about your server; the translation layer changed. Anthropic’s API accepts MCP definitions directly, while OpenAI receives them converted into function-calling format. Ambiguous tool names and nested arguments survive that conversion less well. Run both against the same server and compare which tool each picks.
Test Your MCP Server With ChatGPT & GPT-5 — Free, No API Key | MCP Playground