One email each morning — yesterday's AI, sortedGet it in your inbox

METAL LAB

nvidia/skillevaluator

137PythonApache-2.0

A three-stage checkup for the 'skill folders' that plug new abilities into AI agents

SkillEvaluator is an open-source tool that checks agent skills, which are folders of instructions and files that extend AI agents. It runs a validation stage for safety and formatting, a deduplication stage to catch overlap with existing skills, and a live evaluation stage that tests whether a skill actually helps an agent perform. Built by NVIDIA under Apache 2.0, it's currently labeled as experimental, community-supported software.

What it does

  1. A skill is a folder containing a SKILL.md file plus supporting materials that extend an AI agent, following the Agent Skills specification
  2. Tier 1 validation checks schema, personal data (PII), licensing, quality, and security issues, and several of these checks run without any API key
  3. Tier 2 deduplication uses embeddings, a technique that turns text into numeric vectors, to measure how much a new skill overlaps with existing ones
  4. Tier 3 live evaluation runs the skill with a real agent inside a Docker, local, or cloud sandbox to see how it changes agent behavior, and can auto-generate its own test dataset
  5. Each tier can be run independently, and the tool integrates with other open-source projects like NVIDIA's SkillSpector security scanner and the Harbor agent evaluation framework

Why it matters

As teams start attaching arbitrary skill folders to AI agents, there's a real risk of hidden security flaws, leaked personal data, or redundant instructions that hurt performance, so a standardized pre-deployment check matters. This gives skill authors and agent-pipeline teams a concrete way to vet quality and safety before shipping.

Terms in this repo

  • Agent Skill · A folder of instructions and files, including a SKILL.md, that extends what an AI agent can do
  • Quality gate · A checkpoint that blocks progress unless certain criteria are met
  • Embedding · A technique that converts text meaning into numeric vectors so similarity can be measured
  • Sandbox · An isolated environment for running code safely without affecting the real system
  • PII scan · A check for personally identifiable information, like names or addresses, hidden in a document

Repository description (English)

Multi-tier framework for evaluating AI agent skills with quality gates, semantic overlap detection, synthetic evaluation dataset generation, and live agent evaluation that measures how skills affect agent behavior.

Open on GitHubProject site

Coverage

Trending repos

All repos →

Latest from METAL LAB