What the three levels mean. Risk level is a single-tag summary of a capability's blast radius. Low means read-only or scoped, sandbox-friendly. Medium means it writes files, calls external APIs with limited tokens, or drives a browser. High means shell access, database writes, wide file-system reach, or sensitive credential handling.
How the level is set. Each capability's permission scope — shell, write, token, database, browser, local files, network — is mapped to a severity bucket, and the highest bucket wins. Sensitive tokens push medium up toward high depending on the token's authority. The bucketing is deterministic; you can see the individual permission axes on the detail page and reconstruct the tag yourself.
Why it matters for adoption. Risk level tells you what needs to happen before you can safely try the capability. A low-risk tool can go on your daily driver machine. A medium-risk tool wants a scratch project. A high-risk tool wants a sandbox, revocable credentials, and a review of exactly what it will touch. Use it to plan the trial, not to skip it.