01
Who's out there?
Find every AI agent you're running
- See every agent, who owns it, what it can touch
- Catch agents nobody told you about
- Works with Claude Desktop, Cursor, MCP, and your own tools
- One line to install
Free & open source
Your AI can do
a thousand things.
Should it?
AI agents now read your files, move your money, and email your customers. One wrong move and your data is gone. NIBVOK stands in the doorway, checks every single action, and stops the bad ones before they happen.
WORKS WITH
MCP serversOpenClawClaude Desktop
CursorYour own toolsOpenAI-compatible models
It's not just for engineers
AI is in your home too.
The same protection that guards a company's data guards a family's chat. NIBVOK SafeChat is our consumer side — the same gate, watching over the conversations your kids actually have.
- Flags anything that shouldn't be said to a child
- No data sold, no profiles built
- Built on the same open gateway
The rules are just words
You write the rules in plain language. NIBVOK enforces them.
No cryptic config files. You describe what your agent is allowed to do, and the gateway checks every action against it — before it runs, not after.
- Reads like a policy, not a program
- Changes take effect immediately
- Every rule is versioned and auditable
4 rules · 1 agent
active ✓
How it works
Four questions. Every single time.
Think of NIBVOK as a security guard at the door. Every request your AI makes gets checked before it walks through.
02
What's it allowed to do?
Set the rules, once
- Say yes to specific tools, no to everything else
- Protect customer data and passwords by default
- Make big or permanent actions wait for a human
- Cap what it can spend, delete, or send
03
What's it doing right now?
Watch it live
- See every action as it happens
- Block data leaks and sneaky prompt attacks
- Get an alert when an agent acts out of character
- Full history of every request
04
What do I do about it?
Decide once, apply everywhere
- Shut an agent down mid-task
- Take away access instantly
- Replay exactly what happened
- Hand auditors real proof, not screenshots
The gate
Three answers. Nothing in between.
Every request gets one of three answers, in milliseconds. Your AI never gets to decide for itself whether it's allowed.
ALLOW12ms
Go ahead
This action is on the approved list. It runs, and we write down exactly what happened.
CONFIRMwaiting
Ask a human first
Big or permanent actions pause here. A person says yes or no, and that answer is recorded too.
DENY8ms
Not happening
Off-limits, leaking data, or tricked by an attacker. It never reaches the tool, and your AI gets told why.
Everything goes through the door.
Nothing gets past unchecked. Not a file read, not an email, not a payment.
This already happened
Real disasters a gate would have stopped.
These aren't hypotheticals. Each one is a documented case where an automated system did something nobody approved — and nobody stopped.
DENY
Zillow's home-buying AI
Bought $420 million of houses it had overvalued, with no limit to stop it. A spending cap would have ended it.
DENY
Tesla Autopilot
Switched itself on for roads it was never built for. A rule about where it may operate would have refused.
OVERRIDE
Uber's self-driving car
Saw a pedestrian six seconds early, but braking was switched off by design. An outside monitor would have braked.
CONFIRM
Air Canada's chatbot
Made up a refund policy, and the airline was held to it. A check against the real policy would have caught it.
The record
A paper trail your AI can't erase.
Every yes, every no, every "ask a human" gets locked into a chain that can't be edited or faked. If something goes wrong, you can prove exactly what happened — and so can your auditor, without taking our word for it.
- Nothing can be changed after the fact
- Replay any session from start to finish
- Anyone can verify it independently
- Built for auditors and lawyers, not just engineers
entryanswerfingerprint
#1041ALLOW9f2a…c41d
#1042DENY7b3e…88f0
#1043CONFIRMc14d…2a97
#1044ALLOWe0f5…6b12
chain intact
verified ✓
The whitepaper
Download the whitepaper
Read it over a coffee.
The full technical case for runtime governance — why checking AI after the fact doesn't work, what a gate actually looks like, and how the record is built so nobody has to trust us.
MIT-0free to use, forever
R-1…R-22our own rules, in public
SHA-256every decision fingerprinted
Put a guard at the door.
Install the open gateway in minutes, or have us walk you through a live demo.