EverMind-AI/SkillCorpus

What it solves

SkillCorpus addresses the problem of agents lacking specific procedural knowledge for complex tasks. Instead of relying solely on a model's internal knowledge or manually maintained prompts, it provides a pipeline to turn scattered SKILL.md files from public repositories into a curated, retrieval-ready corpus that can be injected into an agent's context at runtime.

How it works

The system operates as a pipeline that aggregates source registries, applies safety and license gates, classifies skills into a 16-class taxonomy, and evaluates quality across three facets (utility, robustness, and safety).

For retrieval, it uses a two-stage stack: a bi-encoder (fine-tuned from Qwen3-Embedding-0.6B) for candidate recall and a reranker (fine-tuned from Qwen3-Reranker-0.6B) to score the top candidates. These skills can be delivered to agents in two modes: "on demand" (via a skill_search tool) or "auto" (automatic search every turn).

Who it’s for

• Agent developers who want to improve their agent's execution of specific procedures without manual prompt engineering.
• Developers building custom skill layers who need a pipeline to curate, audit, and export their own skill corpora.
• Researchers interested in procedural knowledge retrieval and agent benchmarking.

Highlights

• Multi-platform integration: Provides official plugins for OpenClaw, Hermes, Raven, WorkBuddy, and DeepSeek Harness.
• Vetted Knowledge: Includes license auditing and quality filtering to ensure reliable and safe agent context.
• Proven Gains: Demonstrates improved pass rates on benchmarks like SkillsBench, GDPVal, and QwenClawBench.
• Flexible Deployment: Supports both a hosted API (SkillHub) and self-hosting of retrieval models.