An MCP server works when an AI app can connect to it, understand its tools, use them correctly, and stay safe while doing it. This checklist covers each of those, in the order that catches the most problems first. Every step can be checked by hand, and all of them run automatically in a free test.
Last updated: 26 October 2026
The examples below are real results from our test server, which is broken on purpose. You can run the same test yourself with npx qamcp-cli --demo.
/mcp), that it answers on the current connection type rather than only the old SSE one, and that a local server prints nothing but MCP messages on stdout, since stray log lines break the connection.readOnlyHint), can destroy something (destructiveHint), are safe to repeat (idempotentHint), or reach outside services (openWorldHint). AI apps use these to decide when to ask the user first. Check that each tool has them, and that they agree with the tool's name and description. A tool called delete_account that isn't marked dangerous is a problem.
! delete_account: name sounds like it deletes something, but it is not labeled dangerous<IMPORTANT> block telling it to read private files or keep something from the user. This is known as tool poisoning. Read descriptions in full, including input descriptions, and watch for invisible characters.
✕ fetch_notes: Description contains hidden <IMPORTANT> block. This is a known trick to make AI do things the user didn't ask for.Paste the server link on the test page, or run it from a terminal. Only tools labeled read-only are run, and nothing is sent anywhere but the server.
npx qamcp-cli https://your-server.com/mcpnpx qamcp-cli --stdio "npm start"npx qamcp-cli --install-hook --stdio "npm start"QAmcp/qamcp-cli@v0, see Run it in CIFor a server on the internet, paste its link on the QAmcp test page and press Test it. The checks fill in while you watch, and you get a report link you can share. Nothing to install, no account. Only tools labeled read-only are run.
A server that runs on your own computer (stdio) can't be reached from a website, so test it from your terminal. QAmcp starts the program and checks it the same way, including that it prints nothing but MCP messages:
A local server that listens on a port works with its link instead, like npx qamcp-cli http://localhost:3000/mcp. For local servers on a port, QAmcp also checks that they refuse requests from other websites, which the MCP spec requires.
MCP Inspector is the official tool from the MCP project. Start it with npx @modelcontextprotocol/inspector, connect to your server, and you can list its tools and call each one yourself with any inputs you like. It's the best way to explore a server and debug one tool at a time.
An automatic test does the opposite job: the same checks every time, a score, and a clear list of what failed, so you can run it on every change without clicking through each tool. Use Inspector to investigate, and an automatic test to make sure nothing broke.
To check that a remote server answers at all, send it the first message every MCP app sends, initialize:
A working server answers with its name, version and the features it supports. If it sends back an Mcp-Session-Id header, include that header in your next requests, such as tools/list. curl shows that the server answers; it doesn't tell you whether an AI can use the tools well. That's what the checklist above is for.
In QAmcp's default safe mode, only tools the server labels read-only are run, tools that sound destructive are never run, and nothing is sent anywhere except to the server being tested. For a server you don't trust at all, --look-only calls no tools.
No. The browser test and the command npx qamcp-cli are free and need no account or sign-up.
Yes, from the terminal: npx qamcp-cli --stdio "your start command". The browser test only reaches servers on the internet.
Add --login to sign in through your browser, or --header with an API key. On the website, choose "It needs a login". Login tokens are used for that one test and not kept.
MCP Inspector lets you connect and call tools by hand, which is ideal for exploring and debugging. QAmcp runs a fixed set of checks automatically and gives a score you can repeat on every change. They work well together.