Concepts
Trust review
How the SkillSpector trust scanner checks every imported Skill, what each result means, and what package data the full review sends to Azure OpenAI.
- Audience
- Anyone importing Skills, and owners and admins deciding what their workspace can trust.
- Updated
- 2026-10-25
Every Skill you import is checked by the SkillSpector trust scanner before Fygment saves it. You see one of three results: Safe, Caution or Do not install. A running scan is waited for; a scan that fails after retries imports nothing. The scanner lowers risk, but it cannot guarantee that a Skill is safe.
When the review runs
Fygment runs the trust review on imports from the web app, from AI tools connected through MCP, and from the signed-in CLI.
It runs again when a version is approved, using the package Fygment stored, and when a proposal that changes the package is approved.
On the import page you see the findings for each Skill before you import it.
For a GitHub import, a quick scan runs first. A quick Safe goes ahead. A quick Caution, or a quick scan that failed or left a file unread, gets the full review with a language model. The model can raise a result but never lower it.
What each result means
SAFE: the import goes ahead.
CAUTION: the findings are shown in plain language and the Skill waits for a person to say yes. Rejecting leaves it unpublished. A Caution caused only by a file the Skill mentions but does not include is published with a short note instead.
DO_NOT_INSTALL: the import is refused and nothing is saved. A language-model review never lowers a result.
Still checking: if the scan is still running after about 45 seconds, nothing is imported yet. Ask again later; the scan keeps going and you get the same one back.
Could not check: Fygment retries a failed scan according to why it failed. If every attempt fails, nothing is imported and there is no way to override it.
A file the scanner could not read: the Skill waits for a workspace owner or admin to say yes. A member who hits this is told the Skill was not imported and that an owner or admin must import it.
Trust review from an AI tool or the terminal
skill_scan_trustMCP tool. Scans a Skill stored in Fygment, or a package before you import it, and returns a summary of the findings.
skill_approve_importMCP tool. Approve or reject a version the trust review is holding, after the findings have been explained. Owners and admins decide; a member can accept a Caution only on their own personal Skill. The package is scanned again before it publishes.
fygment scan --path=<dir>CLI. Runs a quick static scan of a local Skill folder before you import it. It needs the SkillSpector scanner installed on your machine.
Privacy
The full review uses Azure OpenAI
A full trust review sends the package contents to Azure OpenAI for analysis. Fygment keeps only a short summary of the result, not the package contents.
For more on what Fygment stores and shows, see Security and privacy boundaries at /docs/security-and-privacy.
Limits
A clean result is not a guarantee
The scanner reduces risk. It does not prove that a Skill is safe. Read what a Skill does before you share it with your team.
Related docs
- Import Skills
Import turns a Skill you already have into a Fygment Skill. Each import is validated and trust reviewed. A Skill that passes is ready for you to use right away. Sharing it more widely is a separate step.
- Review and approve Skills
Members can use the Skills they import right away. Review is for sharing more widely and for new versions. Each decision is made by someone with authority over that Skill or team, and every version approval re-runs the trust review.
- Security and privacy boundaries
The Fygment trust model is metadata-minimizing: keep customer Skill content, package bodies, prompts, secrets, tokens, and raw local paths out of admin, support, analytics, and billing surfaces.
- CLI reference
The CLI is the local bridge between a governed workspace and writable AI harness targets. Use it for setup, install state, update checks, setup checks, and selected imports.