Log inSubmit server

MCP Protocol Inspector & Response Debugger

Connect to a live MCP server to list and call its tools, or validate Model Context Protocol (MCP) JSON-RPC 2.0 payloads, verify content block schemas, catch protocol errors, and preview how AI clients display execution results.

  • Free
  • No sign-in
  • Live calls proxied server-side

1 Connect

Any public Streamable HTTP endpoint. Requests are proxied server-side, so CORS never blocks you.

2 Call a tool

Not connected

Connect to a server to list its tools, then call one here.

Understanding Model Context Protocol JSON-RPC specifications

The Model Context Protocol communicates over standard JSON-RPC 2.0 transports (stdio pipes or Server-Sent Events). When an AI client like Claude Desktop, Cursor, or Windsurf calls a tool or queries a resource, it expects responses to adhere strictly to protocol schemas.

Common integration issues — such as missing content arrays, unhandled base64 image strings, or omitted mimeType headers — can cause AI clients to silently drop tool outputs or fail turns. Use this diagnostic inspector during development to verify payload compliance instantly.

Frequently asked questions

Paste any JSON-RPC request or tool execution response payload into the inspector above. The diagnostic engine checks specification compliance (jsonrpc: "2.0", content array structures, MIME types, isError flags) and renders a live visual preview of how AI clients display the result.

A valid MCP tool call response must contain a content array of objects. Each content object requires a type property ("text", "image", or "resource"). Text objects must have a string text property, and image objects must specify a valid mimeType (e.g. "image/png") and base64-encoded data string.

When a tool returns an image content block with mimeType and base64 data, Claude Desktop and Cursor render the image inline within the conversation turn. Resource content blocks with URIs allow clients to read text or binary files directly.

Setting isError: true inside a tool call result tells the AI model that the tool execution failed (e.g. network error, invalid SQL query, or missing file), allowing the model to analyze the error output and adjust its strategy.
How to build an MCP serverValidate config files