Your MCP server builds but the host can't see its tools. What's the first debugging move?

Attempt it

Run mcp-inspector, the official browser-based debugging UI for MCP servers. Launch it with npx @modelcontextprotocol/inspector <server-command> (for example, npx @modelcontextprotocol/inspector npx @modelcontextprotocol/server-filesystem /tmp). The inspector spawns your server, connects to it as a full MCP client, and opens a local web UI where you can click 'list tools' / 'list resources' / 'list prompts', manually invoke any tool through a form-based interface, and see the raw JSON-RPC traffic with full error context. About 80% of 'server does not show up in the host' bugs are visible in the inspector within a minute: wrong command path, missing env var, schema validation failure, or the server crashing on startup. Reach for it before restarting Claude Desktop or Cursor.

Your answer

**mcp-inspector** is the official browser-based debugging UI for MCP servers and the tool you should reach for first whenever a server does not appear in Claude Desktop or Cursor. Launch it with one command: ``` npx @modelcontextprotocol/inspector npx @modelcontextprotocol/server-filesystem /tmp ``` The inspector spawns the server you pointed it at (here, the filesystem reference server scoped to `/tmp`), connects as an MCP client, and opens a local web UI in your browser. From that UI you can: - See the full JSON-RPC traffic between the inspector and your server in real time. - Click 'list tools' to see what your server advertises, with the JSON Schema for each tool. - Click any tool, fill out arguments in a generated form, and invoke it. - See the response or error with full context, formatted cleanly. Why this matters: when a server 'does not show up' in Claude Desktop, the cause could be a wrong command path, a missing env var, a schema bug, or a startup crash. Restarting the host teaches you nothing. The inspector surfaces the actual error within seconds. Reach for it any time you are building, wiring, or debugging an MCP server. It is by far the fastest feedback loop the ecosystem offers.

**mcp-inspector** is the official browser-based MCP debugger, maintained by the MCP team and shipped on npm as `@modelcontextprotocol/inspector`. It is the single most useful development tool in the MCP ecosystem in 2026. The launch pattern is one command: `npx @modelcontextprotocol/inspector <your-server-command>`. For a Python server at `./server.py`, that is `npx @modelcontextprotocol/inspector python ./server.py`. For a remote server, use `npx @modelcontextprotocol/inspector --url https://example.com/mcp`. The inspector acts as a full MCP client, so it drives any MCP-compliant server without the server needing to know about the inspector. The web UI gives you five surfaces in one page: - A live JSON-RPC traffic log with timestamped requests and responses. - A tools panel: click `tools/list`, see what the server advertised with each tool's JSON Schema. - Per-tool invocation forms: the schema is rendered as input fields; fill them, click 'call', see the structured response or error. - Equivalent panels for `resources/list` / `resources/read` and `prompts/list` / `prompts/get`. - A stderr stream from the spawned server, catching console output that hosts normally swallow. When to reach for it: any time a server fails to appear in a host. About 80% of those bugs surface in the inspector within a minute, typically one of: wrong command path, missing or mistyped env var, schema validation failure on a tool definition, or the server crashing on startup. The inspector shows all of these as a single visible error; the host just shows nothing.

Quick reference

TL;DR

Run `npx @modelcontextprotocol/inspector <server-command>` to launch the official browser-based MCP debugger, which acts as a full MCP client and lets you list and invoke tools through a web UI.

Memory aid

Sign in for the mnemonic that makes this stick. See Pro pricing.

Easy to grasp

Imagine you built a small radio that is supposed to broadcast to a specific app, but the app says it cannot hear you. Instead of guessing what went wrong, you want a separate test receiver that listens on the same frequency, shows you exactly what the radio is sending, and lets you press each button one at a time to see what happens. mcp-inspector is that test receiver for MCP servers. You point it at your server, it shows every message in plain view, you click 'list tools' and 'call this tool with these arguments' on a web page, and any error shows up in readable format instead of buried in a host log.

Key concepts

mcp-inspector MCP debugging JSON-RPC tracing

Detailed answer & concept explanation

~6 min read

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example. Click to expand.

Everything you need to truly understand this topic: intuition, mechanics, step by step explanation, code, formulas, and worked example.

Sign in for 5 free unlocks/day or go Pro for unlimited. See Pro pricing.

Quick scan · 2–4 min

Everything important, quickly.

3 min: the npx launch command, the inspector as a full MCP client, the five UI panels (traffic, tools with forms, resources, prompts, stderr), the four-step triage for missing from host bugs, and where the inspector falls short (host-specific quirks, CI).

Real-world examples

Real products, models, and research that use this idea.

A junior writes their first FastMCP server in Python, runs the inspector, sees `tools/list` returning empty because the `@app.tool()` decorator was missing, and fixes it in 30 seconds.
A team debugging a flaky 'resource not found' error uses the inspector to invoke the resource handler with edge-case URIs (trailing slashes, encoded characters) the LLM would never try organically.
Before publishing a community MCP server to npm, the author runs the inspector against the published package to confirm the install and launch experience works end to end on a clean machine.

Interviewer follow-ups

What an interviewer would ask next. Try answering before peeking at the approach.

QWhy is invoking a tool through the inspector's form better than prompting an LLM to call it?

Sign in or go Pro to see the approach for this follow-up.

QWhat is the inspector doing under the hood that a real host also does?

Sign in or go Pro to see the approach for this follow-up.

QCan mcp-inspector debug a remote streamable HTTP server, or only local stdio?

Sign in or go Pro to see the approach for this follow-up.

Don't say this

Red flags and common mistakes that signal junior thinking. Click to expand.

Most common mistake

Restarting Claude Desktop repeatedly to debug a missing server. The inspector connects directly to the server process and surfaces the actual error in seconds, without needing any host running.

Sign in or go Pro to see all red flags and common mistakes. See Pro pricing.

60-second cheat sheet

The night-before-the-interview bullets. Scan these on the way to the call.

Sign in for 5 unlocks/day or go Pro for unlimited revision sheets. See Pro pricing.

References

Primary sources. Skim if you want the original framing.