ccppccpp home

Tools / MCP

MCP Inspector

beta

Enter the URL of a remote MCP server. The inspector runs the MCP handshake from your browser, lists the server's tools, resources and prompts, and shows each tool the way Claude Code hands it to Claude: its mcp__ name, the description it keeps, the input schema and an estimate of the context it takes. Call a tool from a form built from its schema and read the result as Claude would, with Claude Code's output limits applied.

The Streamable HTTP endpoint (often /mcp) or the legacy SSE endpoint (often /sse).

The name in your Claude Code config. It prefixes every tool: mcp__<name>__<tool>.

Headers

Sent with every request, e.g. Authorization: Bearer <token>. They stay in this tab.

How to use it

  1. Paste the server URL (the /mcp or /sse endpoint) and, if the server needs one, an Authorization header.
  2. Connect: the inspector sends initialize, then lists tools, resources and prompts.
  3. Open What Claude sees to read every tool definition, its token estimate and the problems found.
  4. Pick a tool, fill its arguments and call it to see the result Claude would read.

Questions

Why does the connection fail with a CORS error?
Browsers only let a page read responses from servers that allow its origin. A server meant for desktop clients often does not send Access-Control-Allow-Origin. Add it on the server, or run the server locally with CORS enabled, then connect to http://localhost.
Which transports are supported?
Streamable HTTP (MCP 2025-03-26 and later) and the older HTTP+SSE transport (2024-11-05). Auto tries Streamable HTTP first and falls back to SSE when the server rejects the POST, as the specification recommends. Stdio servers run as local processes and cannot be reached from a web page.
How exact are the token counts?
They are estimates at about four characters per token. Claude's tokenizer is not public, so the real count can differ by a few percent either way.