Docs · Trust
What a badge promises, and what it does not
A skill runs inside a tool-enabled agent with access to your files, shell, and network. That is why nothing here is listed on an automated verdict alone. This page is the short, honest version; the full pipeline lives on the vetting page.
The one rule
Reviewed, never “safe”
One rule governs every trust claim on SkillPort: we say a skill was scanned and reviewed at version X on date Y. We never call a skill “safe”. A review is evidence about a specific set of bytes at a specific moment, not a warranty about how the skill will behave in future or in your particular environment.
Trust tiers
What each badge means
Community
Imported from an open-source repository under a permissive licence, with attribution. It passed the same scan and review as every other listing, but the identity behind it is the upstream project, not a verified SkillPort seller.
Scanned
The listed version passed the automated static and language-model analysis. The badge names the exact version and date the verdict applies to.
Verified
The flagship tier: scanned and approved by a human moderator, shown as reviewed at version X on date Y. Every new version has to earn the badge again.
Tiers are encoded by shape and label as well as colour, so they stay legible regardless of how you see colour. A badge always refers to a named version, not the skill in the abstract.
Coverage
What the scan looks for
Before a human sees a version, an automated pipeline validates the package, runs deterministic static analysis, and runs a language-model pass. Between them they look for prompt injection (both obvious phrasing and semantic redirection), hidden text such as zero-width characters and right-to-left overrides, dangerous script patterns like piping a download straight to a shell, secret-shaped strings, and suspicious URLs. A critical finding blocks the version; a moderator then reviews what passed and records a written decision bound to the exact checksum.
The full layer-by-layer breakdown, including how the scanner is kept from being talked into a pass and how approval is bound to the package bytes at the database level, is on the vetting page.
Plain limits
What vetting cannot do
Scanning and review reduce risk; they cannot make it zero. Human review of publisher versions, versioned scanners with retroactive re-scans, the ability to yank a bad version, and a site-wide suspension switch all exist precisely because false negatives are possible. Treat any skill, from any marketplace including this one, as code you are choosing to run.
We also do not publish scanner rule details or live finding counts, because that would hand an attacker a mutate-and-resubmit oracle against the detector. Publishers receive sanitised findings only after a human decision.
Accountability
Found something wrong?
Every skill page has a report path that goes straight to the moderation queue. A confirmed harmful listing is suspended immediately, which stops new downloads at the same moment it leaves the catalogue. If you publish, the publishing guide shows how your own versions move through the same checks.