MCP Inspector

The MCP Inspector is the official debugging tool for MCP servers — a browser UI that connects to any server, lists what it exposes, and lets you call tools by hand. When a server misbehaves inside Claude or Cursor, this is where you find out why.

Run it

One command, no install (Node.js ≥ 22.19):

npx @modelcontextprotocol/inspector

That starts the web UI on http://localhost:6274 and prints a launch URL containing a session token — open that exact link. The same package also ships a --cli mode for scripting and a --tui mode for the terminal.

Inspect a local (stdio) server

Append the server's launch command and the Inspector starts it as a subprocess — the same thing your MCP client would do:

# a server you're building
npx @modelcontextprotocol/inspector node build/index.js

# any published server
npx @modelcontextprotocol/inspector npx -y @vendor/mcp-server

# pass env vars with -e; everything after -- goes to the server itself
npx @modelcontextprotocol/inspector -e API_KEY=sk-… -- npx -y @vendor/mcp-server --flag

Click Connect, then work through the tabs: Tools lists every tool with its input schema and lets you call one with arguments you type in; Resources and Prompts do the same for the other two MCP primitives. The notifications pane shows the server's log output — the messages your AI client normally swallows.

Inspect a remote server

Switch the transport dropdown to streamable HTTP (or SSE for older servers) and paste the URL. For servers behind OAuth, the Inspector runs the sign-in flow in your browser and shows each step — which makes it the quickest way to debug a remote server's auth before blaming your client. For token-based servers, add the header instead:

npx @modelcontextprotocol/inspector --cli https://mcp.example.com/mcp \
  --transport http --header "Authorization: Bearer …" --method tools/list

Every remote endpoint on this site's official-server list can be checked this way before you commit it to a client config.

CLI mode

--cli skips the browser entirely: one JSON-RPC method per invocation, results on stdout, exit codes for automation. Useful in CI to catch a server that stopped listing the tools your prompts depend on:

# list tools
npx @modelcontextprotocol/inspector --cli node build/index.js --method tools/list

# call one
npx @modelcontextprotocol/inspector --cli node build/index.js \
  --method tools/call --tool-name mytool --tool-arg key=value

# pipe into jq
npx @modelcontextprotocol/inspector --cli node build/index.js \
  --method tools/list --format json | jq '.tools[].name'

If you already keep servers in a config file, point the Inspector at it with --config path/to/config.json --server myserver instead of retyping commands.

A note on the Inspector's own security

The Inspector's proxy can spawn processes on your machine, so it binds to localhost and requires the session token from its launch URL. Two environment variables loosen that — DANGEROUSLY_OMIT_AUTH=true and DANGEROUSLY_BIND_ALL_INTERFACES=true — and both mean what their names say. Leave them alone unless you're inside a disposable container, and never expose a running Inspector to a network you don't control.

When it doesn't connect

Alternatives

For quick checks you may not need it: Claude Code's /mcp command shows each connected server's status and tools, Cursor's MCP settings page lists tools per server, and a remote server's tools/list can be hit with plain curl if you enjoy hand-writing JSON-RPC. The Inspector earns its place the moment you need to see schemas, call tools with crafted arguments, or watch an OAuth flow fail in slow motion — no other tool shows the whole conversation.

Questions