Open-Source Project Proposes Versioned Evidence Schema for AI Agent Skills

A developer has published an experimental open-source project called Agent Catalog Seed, designed to track verifiable evidence for AI agent skills, plugins, and MCP servers rather than relying on simple popularity scores. The project argues that cataloging a capability is fundamentally different from proving it works for a specific task, separating evidence into distinct stages from static checks to runtime observations. It uses a strict JSON Schema to record provenance, requested permissions, compatibility status, and evaluation results, deliberately marking entries as non-installable and unevaluated until real evidence exists. The schema also captures conflicts of interest, such as whether the evaluator and author are the same person, and warns that metadata alone cannot enforce sandbox security. The author, who maintains the repository, disclosed that AI assistance was used in drafting the article and that all examples in the catalog are synthetic.