Skillet
Browse
⌘K
Log inJoin
BrowseDocs
Log in
docs
  • What is Skillet?
  • Summon a kit
  • Install
  • Add skills
  • Skills & kits
  • Approve updates
  • Publish a skill
  • Teams
  • Safety
  • Privacy
  • FAQ
  • API
  • CLI
  • MCP
  • Scanner
  • Skill.md
  • Versioning
  • What is Skillet?
  • Summon a kit
  • Install
  • Add skills
  • Skills & kits
  • Approve updates
  • Publish a skill
  • Teams
  • Safety
  • Privacy
  • FAQ
  • API
  • CLI
  • MCP
  • Scanner
  • Skill.md
  • Versioning

Safety

How Skillet keeps you in control of other people's skills. You approve every change, and every skill is signed and scanned.

A skill is a set of instructions an AI follows. Running a skill someone else wrote means running their judgment in your tools. Here's how Skillet keeps you in control of that. (For what Skillet knows about you, see Privacy.)

You approve every change from other people

  • Updates from other people wait for you. A skill you added from another author never updates on its own: the new version waits as a plain summary until you approve it. Your own skills and your team's kits update automatically, because you already trust them. You can change either default. See Approve updates.
  • Changes are reversible. Skillet backs up your old version before it writes the new one. If an update goes wrong, run skillet restore and you're back where you were.
  • Skillet only writes where it should. It writes into the folders your AI tools read from, and nowhere else. If a tool is missing, it tells you plainly instead of guessing.

You can tell who published a skill

  • Every published skill is signed by the person who published it, and Skillet verifies that signature on your machine before writing it. A skill tampered with after publishing won't install.
  • Once you've approved a version, that exact version is locked in; Skillet can't fetch an altered copy later.

Scanned, not certified

Published skills are scanned for secrets and obvious injection markers, and flagged ones are held for review. That's scanned, not certified: the scan catches the obvious, but a skilled attacker writing natural-looking prose can slip past. This is why nothing from another author lands without your approval, and why it's worth reading what you install. Every skill page shows its scan (the permissions it uses and anything flagged); see the Scanner reference for what each means.

Capability manifest (what a skill can do)

The skill page also shows a capability manifest: an installer-facing inventory of what the bundle might do (shell commands, network, file writes, and similar). It never blocks install; it informs.

StateWhat you see
Not computedNo panel (scan missing or failed). Never shown as "safe."
Computed, empty (analysis: full)"No capabilities detected": every file was inspected or is a recognized inert shape
Partial analysis"Some files couldn't be analyzed": never presented as inert

An empty manifest only reads as inert when analysis is full. Partial analysis means some executable-shaped content could not be inspected (unknown language, extensionless script, binary, or oversized file).

Common worries

Your worryWhat Skillet does
"Will someone else's update break my setup?"Updates from other people wait for your approval, and your old version is backed up.
"Am I running a stranger's tampered file?"Skills are signed and verified before they're written.
"Will it touch files it shouldn't?"It only writes to your AI tools' skill folders, and stops if something looks off.

Note Scanning and signing help, but the real guardrail is you: you decide what lands from other people, every time. For your own data and what /skillet records, see Privacy.

TeamsPrivacy

On this page

  • You approve every change from other people
  • You can tell who published a skill
  • Scanned, not certified
  • Capability manifest (what a skill can do)
  • Common worries
Skilletseasoned to taste
Get appStatsBlogDocs
    Star3