The scanner that never flies the agent
An Apache static scan of MCP tools, with a fix on the line and no execution.
- Category
- Developer tools
- Who it is for
- Someone shipping an MCP server who has not asked who can delete.
- Possible model
- Open source
- Difficulty
- Medium
- Inspired by
- Checkride · Checkride
- Original launch
- Watch the launch · Nov 2, 2026
Difficulty is an editorial estimate of technical and operational complexity, not a score.
Startup idea analysis by Startup Ideas. Last reviewed Nov 2, 2026. How we read a launch.
The startup idea
Preetham posted Checkride on November 2. Can an AI agent delete files without asking. He built a scanner for MCP servers and agent tools that traces issues to the code and suggests fixes. pip install checkride. The repository is PreethamNoelP/checkride. The license is Apache 2.0. Python 3.11 or later. No dependencies, no account, no network access, and it never runs the code it reads. It finds every function a model can call and flags destructive actions with no human approval in front of them. A check placed after the action does not count. A tool argument named confirm does not count, because the model sets it. Neither does approved = True, or a permission check like is_authorized. MCP elicitation does. Every finding carries a concrete fix. The post says scan and suggest. It does not say the tool never executes the agent, so a flag is not a proof that the delete happened. This page did not run it.
The opportunity is the read before the flight. A smaller product is one repo, no network, and the note that a confirm argument is not a human. A sandbox that actually deletes a file and rolls it back is a different product.
Do not say it ran the agent. Do not count a confirm flag.
The problem
An agent can delete a file because nobody asked.
Who would pay for this?
- Primary
- Someone shipping an MCP server who has not asked who can delete.
- Secondary
- A person who will treat a static finding as a starting point, not a penetration test.
- Early adopter
- One repo, one finding, the no-execution line.
How could it make money?
Possible models. Nothing here is a claim about what Checkride charges.
- No price. Apache 2.0.
- No account.
- The post says pip install.
What would the MVP look like?
- pip install.
- A local scan.
- A fix on the line.
- The never-runs line.
How to validate the idea
- Ask five people whether the demo deleted a file.
- If they say yes, they missed the static limit.
- Do not treat a confirm argument as approval.
How to differentiate
- It reads. It does not run.
- A late check does not count.
- No network.
- Do not become the agent.
Original product launch
This launch is inspiration and evidence that someone shipped, not a partnership. The idea above is our reading of the broader opportunity. Checkride did not write it.
- Product
- Checkride
- Source
- Checkride
- Film
- Checkride, it reads the code and does not run it
- Posted
- Nov 2, 2026 · 0:58
Product that inspired this idea
Preetham's post asks whether an AI agent can delete files without asking, then says he built Checkride to scan MCP servers and agent tools, trace issues to the code, and suggest fixes. pip install checkride. The repository is PreethamNoelP/checkride. The license is Apache 2.0. Python 3.11 or later. No dependencies, no account, no network, and it never runs the code it reads. A check placed after the action does not count. A tool argument named confirm does not count, because the model sets it. MCP elicitation does. Every finding carries a concrete fix. The post says it scans and suggests. It does not say the scanner never executes the agent. This page did not run it.
Posted by @preetham_noel_p. Watch the original launch. No separate product site is listed unless the post itself is the source.
- agents
- security
Related startup ideas
- 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.
- A visible team of agents in a world you can walk into
Use a shared world, not a log file, when several agents need a person to come over and decide.
- An Android remote for coding agents
A phone client for checking projects and steering agents when the laptop is closed.
- An inbox so the agents stay quiet
Collapse every coding-agent thread into one list, and only interrupt when one of them is stuck.
- Coding agents in the menu bar
Show which coding session is working, which is stuck, and let a person reply without opening the app.
- Cut the launch film from the app you already have
A skill that rebuilds the screens in a frame, times them to a track, and renders on the machine the coding agent already uses.
Looking for more ideas?
The full list is startup ideas for 2026, each one tied to a launch you can watch.
Browse startup ideas for 2026