MCP Server Tester

FreeFor developers

Built for MCP server developers. Validate your tool schemas, exercise tools/call with real input, and inspect every JSON-RPC frame your server emits — so you ship a server that actually works in Claude Desktop, Cursor and Claude Code.

Test against real LLMs

See how Claude, GPT and Gemini actually call your tools — schema mistakes show up immediately when a real model runs them.

✦ Free credits on sign-up · no credit card

Loading MCP server tester…

Going to publish your server?

Run the security scanner first — HTTPS, auth, headers, CORS, error verbosity, and 14 more checks.

Run Security Scan →

MCP weekly digest

New servers, security advisories, and tutorials — straight to your inbox. No spam.

Why test your MCP server here?

If you are building an MCP server, the worst time to find a schema bug is when a Claude Desktop user files an issue. This MCP server tester lets you exercise every JSON-RPC method your server implements — initialize, tools/list, tools/call, prompts/get, resources/read, completion/complete, logging/setLevel — and see the raw request and response.

It is what you would write yourself with curl and a JSON pretty-printer, but already built. Point it at http://localhost:3000 during development, or your deployed staging URL before you publish.

A pre-publish testing checklist

  1. initialize responds with protocolVersion, serverInfo, and the capabilities you actually implement
  2. tools/list returns every tool with a complete inputSchema (no any, no missing required fields)
  3. tools/call returns content arrays with the correct MIME types — not bare strings
  4. Tool errors use isError: true with a useful message; do not throw raw exceptions
  5. Schema enums and patterns are honored — try invalid input and confirm a clean rejection
  6. Long-running tools stream progress notifications or finish under your client's timeout
  7. Pagination works on tools/list / resources/list if your server has >50 entries
  8. Auth rejects unauthenticated requests on every protected method (not just initialize)

Common MCP server bugs this catches

  • Schema drift — your TypeScript type says userId: number but the schema you advertise says string
  • Missing required — a field your code assumes is present is actually optional in the schema
  • Bare-string responses — returning "ok" when the spec requires a {type:"text", text:"ok"} content item
  • Stack traces in production — uncaught exceptions leaking file paths to the client
  • Initialize advertising capabilities you didn't implement — clients then call methods that 500
  • Inconsistent IDs — sending a notification response when the client sent a request (or vice versa)

Testing local servers

For HTTP / SSE / Streamable HTTP local servers, paste http://localhost:PORT and connect — this works for most modern frameworks (FastMCP, the TypeScript SDK's StreamableHTTPServerTransport, Python's mcp.server.streamable_http).

For STDIO servers, use the official npx @modelcontextprotocol/inspector tool — browsers cannot spawn child processes. Many developers start STDIO during prototyping and switch to Streamable HTTP for distribution; this server tester then becomes the primary tool.

Frequently Asked Questions

How is this different from the MCP inspector?
Same underlying tool; this page is positioned for MCP server developers (writing servers) versus the inspector (debugging any server). The UI is identical — pick the page that matches how you think about the work.
Can I test my localhost MCP server?
Yes, as long as it speaks HTTP, SSE, or Streamable HTTP. Paste http://localhost:PORT and connect. Most browsers allow this by default; if CORS gets in the way, configure your server to allow the MCP Playground origin during development.
What about testing with a real LLM?
A surprising number of schema bugs only show up when a model actually picks arguments. Agent Studio connects your server to Claude, GPT and Gemini in a single conversation — you can compare how each model interprets your tools and catches things that look fine in the raw tester.
How do I publish my server once it passes?
Read How to Publish Your MCP Server to the Registry — covers the manifest, validation, and submission process.
Should I run a security check before publishing?
Yes — strongly. The MCP security scanner runs 20+ automated checks including auth enforcement, TLS, CORS, and error verbosity. A February 2026 sweep found 8,000+ public MCP servers with no authentication; the scan takes 30 seconds.

Related tools & guides

MCP Server Tester — Validate Your MCP Server Before You Ship (Free) | MCP Playground