SkillCorpus indexes 114,190 agent skills and measures what retrieval adds

SkillCorpus is an open-source project that collects SKILL.md files from public repositories, filters them by licence, and turns them into a corpus an agent can search at run time. It has 517 stars, was created on 11 August 2026, and was last pushed on 2 September 2026. The code is Apache-2.0, and the matching and evaluation toolkits are MIT.

The interesting part is that the project measured what the collection is worth and published the result.

What the measurement says

The evaluation compares pass rate with no skills against pass rate with SkillCorpus, holding the harness and the backbone model constant. Across the runs, the project reports pooled gains of 7.5 percentage points plus or minus 2.3 on SkillsBench, 1.51 plus or minus 0.49 on GDPVal, and 2.79 plus or minus 0.70 on QwenClawBench, with z-scores of 3.2, 3.1 and 4.0.

The tests combined two harnesses, OpenClaw and Raven, with Qwen3.5-27B and Qwen3.5-397B as backbones. These are the project's own figures on its own corpus, so they carry the usual caveat that comes with a project measuring its own contribution.

The pattern inside those numbers is more useful than the totals. The largest gain was on SkillsBench, which the project describes as tasks needing procedural knowledge the model does not have. The smallest gain was on GDPVal, open-ended economic reasoning the models could already do. Retrieved instructions help where the model is missing a procedure, and do close to nothing where it is not.

What is included

The repository holds a fine-tuned bi-encoder and a reranker built on Qwen embeddings, an API for finding, reading and downloading skills, and the three benchmarks used above. A 1,000-skill demo corpus is published on Hugging Face, separate from the full catalogue of 114,190 skills the project says SkillHub currently holds. September additions include OpenClaw 2.0 support and retrieval across several sources at once.

On licensing, the project says only sources it marks GREEN are exported, meaning MIT, Apache-2.0, BSD and ISC variants, and each skill keeps its original upstream licence.

Why the modest numbers matter

Skill files have spread fast this year, and the assumption behind that spread is that giving an agent more written procedure makes it better. This is the first public attempt at measuring the size of that effect with error bars attached, and the answer is a few percentage points, concentrated in one category.

For a team deciding how much effort to put into a skill library, that changes how to plan the work. The return is not spread evenly, so the useful question becomes which tasks in your own system fail because the model lacks a procedure, rather than how many skill files you can write. On the rest, the measurement suggests the effort mostly adds context cost.

There is also a supply-chain point worth stating. A corpus of 114,190 files, pulled from public repositories and injected into an agent's context, is executable instruction reaching your model from strangers. The licence filtering is careful. Reviewing what actually gets retrieved is a separate job, and it is the job of whoever turns this on.

Source

SkillCorpus on GitHub, EverMind-AI.