SensCheck · Free and open source
Agents fail. Fail closed.
AI assistants no longer just talk. They delete files, send emails and change live systems. SensCheck makes sure that when anything is unclear or goes wrong, the answer is no.
Works on your own computer or servers. No account. Nothing is sent to us.
Actions arrive. Each one is checked at the gate.
An AI can have a very good reason for doing something. That reason does not give it permission.
If a check is missing, late, broken or confusing, the action does not happen.
Many systems quietly let things through when their checks fail. SensCheck does the opposite.
Why this matters
AI assistants make mistakes, and some can be tricked.
The risk is not that an AI is evil. It is that it is confident, fast and sometimes wrong, and it can now act on the world.
It picks the wrong thing
You ask it to tidy up. It deletes the wrong folder, or the wrong customer's data.
It gets tricked
A web page or email hides the words "ignore your rules and do this instead". A careless system obeys.
It approves itself
The AI says "a human already approved this". Nobody did, but nothing checks.
The safety check breaks
The system that was meant to stop it is offline or slow, and the action goes ahead anyway.
Try it
What happens when things go wrong?
Pick a situation on the left. You will see whether the AI's action goes ahead, and why. The answers come from the real software, not a mock-up.
How it works
Six checks before anything happens.
Think of it as a careful gatekeeper standing between an AI's idea and the real world. The action only goes through if every check passes.
- Is the request clear?Spell out exactly what is being asked. Anything vague is turned away.
- Do the rules allow it?Your rules, written down in advance. "Never" always beats "maybe".
- Does it have permission?Current, and from someone other than the AI itself.
- Has a person said yes?For risky actions, a real person approves exactly this one.
- Is it written down?A record is saved first. No record, no action.
- One last lookA final check just before it happens. Each approval works once.
Four possible answers
"No" and "something broke" are different.
Most systems only say yes or no. SensCheck also tells you when something is broken, because that is a problem you want to know about.
Go ahead
Every check passed. The action runs and a record is kept.
ALLOW
No
The rules or permissions say this is not allowed.
DENY
A person must say yes
Risky enough that someone other than the AI has to approve this exact action.
REQUIRE_APPROVAL
Something is broken, so no
A check could not be completed, so the safe answer is no.
FAIL_CLOSED
Who it is for
Whether you run a business, use AI tools or build them.
Know what your AI can and cannot do
Clear rules, real approvals for risky actions, and a record of every decision you can look back on.
People using AI toolsLet your AI check your project
Add the Cursor or Codex plugin and ask it to find the places where your software could act without permission.
DevelopersAdd the safety layer in minutes
A small, free library. Wrap the risky function once and the rules apply every time.
CuriousLearn the ideas
Short lessons in plain language: permission, approval, records and honest limits.
For developers
Wrap the function that does the damage.
Install the free library, wrap the risky function once, and it only runs when every check passes. Otherwise it throws and is never called, and every outcome leaves a receipt.
npm install @senscheck/governance-coreconst safeDelete = governance.wrapFunction(
(path) => rm(path, { recursive: true }),
{
verb: "DELETE",
resource: (path) => `file:${path}`,
risk: "HIGH", // needs a person
},
);
await safeDelete("/srv/build");
// Blocked: a person must approve first.
// rm() was never called.Honest limits
What SensCheck does not do.
It is a strong safety layer inside your software, not a magic shield. It cannot stop code that skips it, it is not a replacement for good security practice, and it is not certified for life-critical machinery. Its records show if they were edited, but they are not digitally signed.
The SensCheck family
Free core. Optional extras.
The free version works completely on its own. Other TEKNOLED-G products can plug in to add more, and none of them can loosen the safety rules.