skillbox
Self-hosted, versioned skills library for AI agents. MCP, scoped clients, and optional Jev recommendations.
Self-hosted, versioned skills library for AI agents. MCP, scoped clients, and optional Jev recommendations.
When recommendations are enabled, Skillbox sends the task and a bounded set of candidate skill IDs and descriptions. Jev returns one Score answer per skill against a five-level relevance rubric (0 through 4); code validates the scores and filters/ranks recommendations.
Rate only candidate skill_0’s usefulness for the task, matching meaning and respecting the task context and skill limitations.
A user task and a representative candidate skill ID and description. The request has one score question per candidate, using the shared five-level relevance rubric.
Retrieve candidate skills for the task and respect catalog bounds.
Send task text and each candidate skill’s ID and description with a shared relevance rubric.
Call the configured provider, including TypeSafe when selected by the user.
Validate each bounded score and return only candidates meeting the relevance cutoff.
Task relevance is evaluated against a versioned skills catalog while provider choice and keys remain user-controlled. Low scores can be filtered instead of returning every catalog match.