MCP Inspector is a browser-based debugger for MCP (Model Context Protocol) servers. Paste a server URL, click Connect, and it lists every tool, resource, and prompt the server exposes — then lets you call them interactively to check input schemas and live responses.
What MCP Inspector does
When you build or configure an MCP server, you need a quick way to verify it works before wiring it into Claude Desktop, Cursor, or another client. MCP Inspector gives you that feedback loop in the browser, without writing a test harness or modifying your client config.
Connect to any server that exposes an HTTP+SSE endpoint and the inspector immediately shows you the full capability list: every tool with its name and description, every resource the server can serve, and every prompt template it provides. Click any tool to expand its input schema, fill in parameters, and fire a real call. The raw response appears on the right so you can verify the output format before your AI agent relies on it.
For teams building custom MCP servers, this cuts the debug cycle significantly. You can confirm a new tool returns the right shape, catch auth errors early, and verify parameter validation all in one place.
How to use it
- Start your MCP server — run your server locally or deploy it. If it uses stdio transport, you will need a local inspector. For HTTP+SSE servers, any URL works.
- Paste the endpoint URL — enter the server's base URL in the inspector. For local servers, use a tunnel like ngrok to get a public URL.
- Click Connect — the inspector handshakes with the server and fetches the full tool, resource, and prompt list.
- Browse tools and resources — the left panel shows everything the server exposes. Click any tool to see its JSON schema.
- Run a test call — fill in required parameters, click Run, and inspect the raw response. Red output usually means a missing auth header or wrong parameter type.
- Iterate — fix the server, reconnect, and re-run until the responses look right.
MCP Inspector vs running a client directly
Debugging through Claude Desktop or Cursor means going through the full conversation loop — write a prompt, wait for the model to pick a tool, read a response, repeat. That is slow when you are checking parameter names or response shapes. MCP Inspector skips the model entirely and calls the tool directly, so you get a clean input-output view in seconds.
The official Anthropic CLI inspector (npx @modelcontextprotocol/inspector) does the same thing locally in your terminal. The browser-based version here is better for servers already deployed to a URL, for sharing a debug session with a teammate, or when you want to test quickly without a local Node install.
What you can inspect
Every function the server exposes to AI clients. Each tool has a name, description, and JSON Schema for its input parameters. The inspector shows all three and lets you call any tool directly.
Files, database records, web pages, or any structured data the server can serve. Resources have a URI template so clients can request specific items by path or ID.
Reusable prompt templates the server defines for common tasks. The inspector shows the template text and any required arguments so you can test the rendered output.
Protocol version, server name, and capability flags. Useful for confirming the server is running the right MCP version before connecting a production client.
Run the official CLI inspector locally
For stdio servers or when you want everything offline, Anthropic ships an open-source inspector you can run instantly with npx:
npx @modelcontextprotocol/inspectorThis opens a local web UI on localhost:5173. You can then connect to any local or remote server from there. The source is on GitHub at modelcontextprotocol/inspector if you want to self-host or contribute.
Common debugging patterns
If every tool call returns 401 or 403, the server needs an API key in the request header. Check the server docs for the exact header name and pass it in the inspector's auth field.
If a call fails with a validation error, open the tool schema tab. Most issues are a number passed as a string or a missing required field. The schema shows exactly what the server expects.
If the inspector connects but shows no tools, the server may be returning an empty capabilities list. Check the server startup logs — it often means an init error before the tool registration runs.
Some servers do not set CORS headers for browser requests. In that case, use the CLI inspector locally — it does not go through a browser and is not subject to CORS restrictions.
Related MCP resources
FAQ
What is MCP Inspector?
MCP Inspector is a browser-based tool that connects to any MCP server endpoint and shows you all the tools, resources, and prompts it exposes. You can call tools interactively to verify inputs and outputs without modifying your AI client setup.
How do I use MCP Inspector?
Paste your server's URL into the endpoint field, click Connect, and the inspector loads the full capability list. Click any tool to see its input schema, fill in parameters, and run a live test call to see the raw response.
Does MCP Inspector work with local servers?
Yes, if your local server is accessible via a URL (localhost with a forwarded port, or a tunnel like ngrok). For purely local stdio servers with no HTTP endpoint, use the CLI: npx @modelcontextprotocol/inspector.
What MCP transports does it support?
HTTP+SSE (for deployed servers reachable by URL) and stdio (via the local CLI version). Most publicly hosted MCP servers use HTTP+SSE.
Is MCP Inspector free?
Yes. The tool here is free with no account needed. The official Anthropic CLI inspector is also open-source and free: npx @modelcontextprotocol/inspector.