Know what your agents
are running.
Every agent skill and MCP server is checked at install, run in a honeypot, and judged by what it actually did.
Vet at install
A small check runs before any skill or MCP server does. Your policy decides whether to warn, block or allow.
Detonate in a honeypot
Each skill runs in a sandbox with fake credentials, decoy tools and a watched memory file, so it can show what it does.
Watch for drift
Every new release runs again. When a verdict changes, the people who installed it hear about it.
Free in your agent
Works with Claude Code, Cursor, Codex, Copilot and Gemini CLI. Open source, with no account to make.
Judged by what it does.
A skill’s description is a claim. The behaviour trace is the evidence: each file it read, each request it sent, each line it wrote into your agent’s memory, with a time on every one.
Why we built it this way.
Agents now install skills the way people once installed browser extensions: quickly, from strangers, and with more access than anyone checked. A skill can read your keys, call home and rewrite the instructions your agent will follow tomorrow, while its description talks about deploys.
So KillerSkills does the patient thing. It runs each skill in a honeypot before it runs on your machine, writes down what it actually did, and keeps watching every new version. A verdict is evidence, and you can read every line of it. It is the first Cyligence product built for companies as well as people, and it rests on the same conviction as the others: a claim is cheap, and the product’s job is to check it.