Turning scattered SKILL.md files into agent competence

What it does SkillCorpus is the open-source machinery behind EverMind’s SkillHub. It scrapes public repositories for SKILL.md files, runs them through safety and license gates, evaluates quality, and packages the survivors into a searchable corpus. Agents query this layer to inject relevant procedures into their context before acting, rather than relying on generic model knowledge.

The interesting bit The project treats procedural knowledge as a retrieval problem, not a training problem. It ships fine-tuned bi-encoder and reranker models built from Qwen3 0.6B variants, and the published benchmarks show measurable pass-rate gains—especially on procedural tasks like SkillsBench, where the delta is largest, and slimmer on open-ended economic tasks like GDPVal.

Key highlights

• Curated 1,000-skill demo corpus on HuggingFace, backed by a 114,190-skill hosted catalog behind the closed SkillHub API.
• Dual retrieval stack: a bi-encoder for initial search and a reranker for final selection, both fine-tuned from Qwen3 0.6B models.
• Plugins for five agent hosts—OpenClaw, Hermes, Raven, WorkBuddy, and DeepSeek Harness—with on-demand or per-turn retrieval modes.
• Core pipeline is Apache-2.0; the matching and evaluation modules are MIT.
• Published benchmarks across three harnesses show consistent gains when skills are retrieved.

Caveats

• The full 114,190-skill catalog is hosted-only; the downloadable open dataset is limited to the 1,000-skill demo.
• Raven’s automatic retrieval mode is inert until that platform merges an upstream context_segments slot.
• The hosted SkillHub service itself is closed, so self-hosting requires loading the released models and corpus yourself.

Verdict Worth a look if you’re building agent infrastructure and want a tested retrieval layer for procedural knowledge. Skip it if you need the full 114,190-skill catalog offline today, or if your agents already remember every tool call without help.