MCP playground.

Call the thinkhard.si MCP server from this page and see exactly what goes over the wire: the JSON-RPC request, the answer your AI would read, and the same call as curl, JavaScript or Python, ready to copy.

Request

Call

Calls from this page go to https://thinkhard.si/mcp like any other client, and count toward the limit of 600 an hour. Nothing you send is stored.

Response

Press Send to call the server.

The same call in code

A public MCP server for testing your client

Building an MCP client, an agent framework or a gateway? https://thinkhard.si/mcp is a real server you can test against with no key: unlike an echo demo, it has tools with input and output schemas, 300+ prompts with arguments, and answers that stay the same from one run to the next.

Tools: five read-only tools, each with an outputSchema and structuredContent, plus isError answers you can trigger on purpose, such as grade_prompt with an empty prompt.
Prompts: prompts/list pages 100 at a time with nextCursor, so it exercises your paging, and prompts/get fills arguments.
Protocol: Streamable HTTP with JSON answers, stateless, protocol versions 2024-11-05 to 2025-11-25, notifications answered with 202, CORS open for browser clients.
Errors: an unknown tool gives -32602, an unknown method -32601, a bad body -32700. The full list is in the tools reference.

Please keep automated tests under 600 calls an hour. For heavy load testing, run the local server instead.

What this shows

The server is stateless: every call stands on its own, with no session to open first. That is why each request here is a single POST.
“What the AI reads” is the text an assistant gets back. structuredContent holds the same answer as data, for code, and matches the tool’s output schema in the tools reference.
check_robot_speak and fact_check_list are missing here because they run only in the local server, so your drafts stay on your computer. On the website, they are the Robot-Speak Checker and the Fact Check List.