MemTensor/skills-vote

What it solves

SkillsVote addresses the challenge of managing large-scale agent skill libraries. As the number of available skills (e.g., SKILL.md files) grows into the millions, agents cannot simply load a static list. They need a way to determine which specific skills are relevant to a given task, verify if those skills actually helped, and evolve the library by updating or creating new skills based on real-world execution feedback.

How it works

SkillsVote implements a lifecycle governance loop for agent skills:

1. Collection and Profiling: It gathers skills from open-source or private libraries and analyzes their requirements, dependencies, and quality.
2. Just-in-Time Recommendation: Instead of loading all skills, it recommends relevant skills to the agent immediately before task execution.
3. Trajectory-Based Attribution: After a task is completed, it uses execution trajectories and verifier signals to attribute the outcome to specific skill usage.
4. Feedback-Driven Evolution: It uses the attributed feedback to update existing skills or generate new reusable skills, preventing the accumulation of noisy data.

Who it’s for

Developers building coding, research, or workflow agents (such as Codex, Claude Code, or OpenClaw) who need to manage a vast library of reusable execution capabilities.

Highlights

• Massive Skill Library: Includes over 1.68 million discovered skills, with 790K verified via the Anthropic skill validator.
• Flexible Integration: Offers both a hosted cloud service for recommendation and feedback, and a skills-vote-local version for private libraries.
• Proven Performance: Demonstrated improvements on agentic coding and terminal benchmarks like SWE-Bench Pro and Terminal-Bench.
• Lifecycle Management: Moves beyond simple retrieval to a full loop of recommendation, attribution, and evolution.