The MCP test that will not press delete

The startup idea

The problem

Restarting the desktop app is how people find out a server leaks a secret.

Who would pay for this?

Primary

Someone shipping an MCP server who restarts the desktop app to see if it works.

Secondary

A developer who will run Node 20 and point the CLI at a local server.

Early adopter

The demo report, then one real server, with writes left unrun.

How could it make money?

Possible models. Nothing here is a claim about what QAmcp charges.

• No price. MIT.
• npm is the install.
• The post does not name a paid tier.

What would the MVP look like?

1. One URL or one local stdio server.
2. A hidden-instruction finding.
3. A leaked secret in an error, named.
4. Writes marked not run.

How to validate the idea

• Ask five server authors to run the demo before their own URL.
• If a write tool ran, safe mode failed.
• Do not time the film and call it the test.

How to differentiate

• Results are pass or fail, not a model score.
• It talks only to the server you name.
• Mutating tools stay off.
• Do not become the MCP host.

Product that inspired this idea

Abhi's post says one command tests an MCP server in 12 seconds. Tools, bad inputs, prompt injections, and leaked secrets. Free and open source. The readme says MIT, Node.js 20 or newer. It connects like an AI app and sends nothing except to the server you name. The demo starts a server they broke on purpose and takes about 30 seconds. It flags a hidden instruction that tells the AI to read an SSH key, and an error that shows a database password. Tools that change data are not run, ever. A read-only tool with no safety label is skipped in safe mode and runs with --owner. The post's 12 seconds is the length of the film. The readme's demo is about 30. This page did not test a server.

Posted by @AbhiRam2371650. Watch the original launch. No separate product site is listed unless the post itself is the source.

• mcp
• testing

Related startup ideas

• The suite that stays under the wallpaper

  A MIT Claude plugin that runs SwiftUI tests in the background, and skips the checks that need a focused window.

• A browser inside the coding agent’s terminal

  Open a real browser from the agent session so the agent can see the page it just changed.

• A checklist the coding agent cannot hide

  A plain-language checklist beside Claude Code that pauses for an OK, says when a step is stuck, and puts the limit and the cost on the same screen.

• A coding agent that runs a small model in the browser

  A coding assistant that runs locally in the browser on a small model, for people who will not send the repo out.

• A desktop for making games with the model you already pay for

  A local app that turns a Claude or ChatGPT subscription, or a local model, into a game-making bench.

• A drive that files the document and cites the page

  Upload a pile of files, have them land in a folder a person would recognize, and answer a question with the page it came from.

More developer tools ideas