Skip to content
Back to search

SkillCorpus

Open-source infrastructure that turns scattered SKILL.md files into curated, retrieval-ready agent-skill corpora—with retrieval and evaluation tooling included.

Safety scan

A static text scan, not a guarantee. Select the badge to see what was scanned and every finding.

Checked today

Check it deeper on your machine

It has changed since its scan on 11 Nov 2026.

As a Claude Code plugin, install it and then ask Claude Code:

/plugin marketplace add keithmackay/mackayi/plugin install scanskillsafety@mackayiDo a deep scan of https://github.com/EverMind-AI/SkillCorpus with scanskillsafety

Or from a terminal, clone it and run it:

git clone https://github.com/keithmackay/scanskillsafety-skill ~/.claude/skills/scanskillsafetynode ~/.claude/skills/scanskillsafety/cli.cjs --deep 'https://github.com/EverMind-AI/SkillCorpus'

It checks the exact commit: every file, its packages against OSV (api.osv.dev), how it compares with this page's scan (findsafeskills.com receives the repository's address), and, if you agree, Claude reading every file with your own Claude subscription. It never runs the code.

Badge for your README
findsafeskills safety rating

Markdown

[![findsafeskills safety rating](/api/listings/evermind-ai--skillcorpus/badge.svg)](/listings/evermind-ai--skillcorpus)

HTML

<a href="/listings/evermind-ai--skillcorpus"><img src="/api/listings/evermind-ai--skillcorpus/badge.svg" alt="findsafeskills safety rating" /></a>

Install

i

What was scanned

The files the safety scan read, as written by the repository's author and shown as plain text. Lines with a finding are marked.

README.md · README · 462 lines

View this file on GitHub (read at the repository's current version: the scan did not record which commit it read)

This file has changed since it was scanned, so this is not exactly the text that was scanned, and no line is marked. The findings are listed with the safety scan above.

  1. <!-- SkillHub is the live hosted product; this repository contains the open-source corpus,
  2. retrieval, evaluation, export, and plugin layer behind it. -->
  3. <div align="center" id="readme-top">
  4. <table width="100%" border="1" bordercolor="#d9d9d9" cellspacing="0" cellpadding="0">
  5. <tr><td><img src="https://github.com/user-attachments/assets/2ef7e877-275d-4115-8ddf-f9b49de8ff5d" alt="SkillCorpus banner" width="100%"></td></tr>
  6. </table>
  7. <p align="center">
  8. <a href="https://arxiv.org/abs/2607.15557"><img src="https://img.shields.io/badge/arXiv-2607.15557-b31b1b?labelColor=gray&style=for-the-badge" alt="Paper"></a>
  9. <a href="https://huggingface.co/EverMind-AI"><img src="https://img.shields.io/badge/HuggingFace-EverMind-F5C842?labelColor=gray&style=for-the-badge&logo=huggingface&logoColor=white" alt="Hugging Face"></a>
  10. <a href="https://discord.gg/QHQ2YYp7cH"><img src="https://img.shields.io/badge/dynamic/json?url=https%3A%2F%2Fdiscord.com%2Fapi%2Fv10%2Finvites%2FQHQ2YYp7cH%3Fwith_counts%3Dtrue&query=%24.approximate_presence_count&suffix=%20online&label=Discord&color=404EED&labelColor=gray&style=for-the-badge&logo=discord&logoColor=white" alt="Discord"></a>
  11. <a href="https://github.com/EverMind-AI/EverOS/discussions/67"><img src="https://img.shields.io/badge/WeCom-EverMind_%E7%A4%BE%E5%8C%BA-07C160?labelColor=gray&style=for-the-badge&logo=wechat&logoColor=white" alt="WeCom"></a>
  12. </p>
  13. <p align="center"><strong>English</strong> · <a href="README.zh-CN.md">简体中文</a></p>
  14. </div>
  15. <br>
  16. ## What SkillCorpus gives you
  17. SkillCorpus is EverMind's open-source pipeline for turning scattered `SKILL.md` files from public
  18. repositories into reliable agent context. It aggregates sources, applies safety and license gates,
  19. evaluates quality, and matches task-specific skills before the agent answers.
  20. You can use the live [SkillHub](https://evermind.ai/skillhub) without cloning this repository. Clone
  21. SkillCorpus when you want the open-source machinery behind that experience:
  22. - **Build your own skill layer** — point the pipeline at your own source registry, apply the
  23. curation, safety, and license gates, and export a corpus for your agents.
  24. - **Change the behavior** — modify the taxonomy, quality and dedup rules, retrieval recipe, export
  25. schema, evaluation suites, or host plugins.
  26. - **Keep control of deployment** — self-host the released retrieval models and connect your own
  27. agent host instead of using the hosted SkillHub API.
  28. The core code is Apache-2.0 licensed (`match/`, `evaluate/` and `lifecycle/` are MIT); each skill retains its
  29. upstream license. The public 1,000-skill demo, three agent benchmarks, and live SkillHub show the
  30. result.
  31. https://github.com/user-attachments/assets/4d9a3241-df13-4b20-9798-fb7920069995
  32. <br>
  33. ## &#128293; Latest Updates
  34. - **2026-10-28** The October corpus version holds **300,473 skills**, **178,173** of them under
  35. commercial-safe licenses — same curation pipeline, same schema. Release on HuggingFace **coming soon**.
  36. - **2026-10-23** **SkillRep** — skill retrieval is hard because a request and the skill that answers it are
  37. written in different words, and existing methods each move only one or two dimensions. We give the
  38. **first complete formulation of that representation space** (five dimensions; most prior retrievers
  39. occupy a single point in it), optimize across the whole of it, and reach **state of the art on both
  40. mainstream benchmarks** — a **0.6B** encoder beats the strongest same-size retriever on [SRA-Bench](https://huggingface.co/datasets/WeihangSu/SRA-Bench) by
  41. **+7.02 MRR@10** and leads [SkillsBench](https://huggingface.co/datasets/pipizhao/SkillRouter-Eval-Core). We also make the space **interpretable**: three measurable
  42. quantities show which part of the gap each dimension closes, and when two of them cancel.
  43. **Code and model are open.**
  44. [arXiv:2610.32336](https://arxiv.org/abs/2610.32336) | [code](skillrep) | [HF encoder](https://huggingface.co/EverMind-AI/skillrep-embedding-0.6b) · [reranker](https://huggingface.co/EverMind-AI/skillrep-reranker-0.6b)
  45. - **2026-10-23** **Beyond Gain** — skill hubs keep growing (ours is past 300K) while the models they serve
  46. keep improving, so a hub has to retire, per model, what that model has already internalized. Yet
  47. internalization had no reliable measurement: the standard *retire-on-gain* test reads **identically** for
  48. a skill the model absorbed and one that never worked for anyone. We give the **first reliable measurement
  49. of internalization** — withhold the skill, score the model, and trust a high score only when some other
  50. model does gain from it: under the old rule **more than 90% of [SkillsBench](https://huggingface.co/datasets/pipizhao/SkillRouter-Eval-Core) retirements** rest on
  51. tasks no run ever solved, while ours loses no reward and reproduces **nearly twice as well**. **Code and
  52. data are open.**
  53. [arXiv:2610.32322](https://arxiv.org/abs/2610.32322) | [code](skillcorpus/lifecycle) | [HF 1,251 skills × 4 models](https://huggingface.co/datasets/EverMind-AI/skillcorpus-lifecycle)
  54. - **2026-09-17 · v0.4.0** Adds a **shared skills library** across all six hosts: they read one `~/.evermind-skillsearch/` directory and register their own, so a skill you have in one agent is available in the rest; skills retrieved from a catalog are **kept** instead of discarded at the end of the turn; and changes take effect **on the next turn, with no restart**. Raven and Hermes also gain `skills_dirs` / `SKILLSEARCH_SKILLS_DIRS`.
  55. - **2026-09-02** Adds OpenClaw 2.0 support and smarter skill delivery: retrieve automatically on every query, or let the main agent call `skill_search` on demand.
  56. - **2026-08-27** Supports multi-source retrieval across local skills, EverMind SkillHub, ClawHub, and skillhub.cn, with filtering, deduplication, and final 0–2 selection.
  57. - **2026-08-26** Supports PathGuard placeholder resolution and host-aware paths for skill files and agent workspaces.
  58. - **2026-08-25** Adds official SkillCorpus plugins for WorkBuddy, OpenClaw, Hermes, Raven, and DeepSeek Harness.
  59. <br>
  60. ## Stronger agents, one turn at a time
  61. At answer time, the practical difference is a retrieval layer: SkillHub selects vetted procedural
  62. knowledge for the task and puts it into the agent's context.
  63. <table width="100%">
  64. <tr>
  65. <th>Dimension</th>
  66. <th>Without SkillCorpus</th>
  67. <th>With SkillCorpus</th>
  68. </tr>
  69. <tr>
  70. <td><strong>Context</strong></td>
  71. <td>Model knowledge plus a manually maintained prompt.</td>
  72. <td>Task-specific, license-audited <code>SKILL.md</code> retrieved automatically or on demand.</td>
  73. </tr>
  74. <tr>
  75. <td><strong>Execution</strong></td>
  76. <td>Generic workflows can miss exact steps, edge cases, or supporting scripts.</td>
  77. <td>Procedures, references, and optional scripts arrive before execution.</td>
  78. </tr>
  79. <tr>
  80. <td><strong>Integration</strong></td>
  81. <td>Each host maintains its own collection of task instructions.</td>
  82. <td>One curated skill layer serves OpenClaw, Hermes, Raven, WorkBuddy, DeepSeek Harness, and other hosts.</td>
  83. </tr>
  84. </table>
  85. The result is the same agent with better task-specific procedures available at the moment it needs
  86. them — stronger execution without asking users to memorise skill names or wire up tool calls.
  87. <br>
  88. ## Results
  89. Pass rate with no skills → with SkillCorpus, same harness, same backbone
  90. ([paper, Table 1](https://arxiv.org/abs/2607.15557)):
  91. | Harness × backbone | SkillsBench | GDPVal | QwenClawBench |
  92. |---|---|---|---|
  93. | OpenClaw × Qwen3.5-27B | 8.8 → **13.0** | 81.2 → **83.1** | 65.2 → **66.7** |
  94. | OpenClaw × Qwen3.5-397B | 11.1 → **16.9** | 82.2 → **84.0** | 65.7 → **67.0** |
  95. | Raven × Qwen3.5-27B | 10.0 → **16.5** | 82.6 → **83.8** | 66.9 → **70.8** |
  96. | Raven × Qwen3.5-397B | 9.2 → **22.6** | 84.0 → **85.2** | 68.8 → **73.2** |
  97. | **Pooled ∆** | **+7.5**±2.3 (z=3.2) | **+1.51**±0.49 (z=3.1) | **+2.79**±0.70 (z=4.0) |
  98. The gain is largest where the task needs procedural knowledge the model does not already have
  99. (SkillsBench), and smallest on open-ended economic tasks it can already do (GDPVal).
  100. <br>
  101. ## SkillHub integrations
  102. SkillHub brings skill retrieval to the five agent platforms below. Choose a platform to open its
  103. plugin guide:
  104. <table width="100%">
  105. <tr>
  106. <td width="400" align="center"><a href="skillcorpus_plugin/engine-typescript/README.md"><img src="https://avatars.githubusercontent.com/u/148330874?s=200&amp;v=4" alt="DeepSeek Harness" width="72"><br><strong>DeepSeek Harness</strong></a></td>
  107. <td width="400" align="center"><a href="skillcorpus_plugin/plugin-hermes/README.md"><img src="https://github.com/user-attachments/assets/477eebc4-e615-4425-921e-368d7667e491" alt="Hermes" width="72"><br><strong>Hermes</strong></a></td>
  108. <td width="400" align="center"><img src="https://github.com/user-attachments/assets/01d948fe-1e2b-48e8-9b32-b8057cb3f336" alt="OpenClaw" width="72"><br><strong>OpenClaw</strong><br><a href="skillcorpus_plugin/plugin-openclaw/README.md">1.x</a> / <a href="skillcorpus_plugin/plugin-openclaw2/README.md">2.0</a></td>
  109. <td width="400" align="center"><a href="skillcorpus_plugin/plugin-raven/README.md"><img src="https://github.com/user-attachments/assets/27e1ea63-69d4-48b3-a884-7f0355926907" alt="Raven" width="72"><br><strong>Raven</strong></a></td>
  110. <td width="400" align="center"><a href="skillcorpus_plugin/plugin-workbuddy/README.md"><img src="https://github.com/user-attachments/assets/ab2157dc-90fc-4196-bbf3-87066820f7b4" alt="WorkBuddy" width="72"><br><strong>WorkBuddy</strong></a></td>
  111. </tr>
  112. </table>
  113. Two modes, one setting. **On demand** — the default — gives the agent a `skill_search` tool and
  114. lets it decide: a long task pays for retrieval at the step that needs it and nothing on the turns
  115. that do not. **`mode: auto`** is the older behaviour: search every turn, before the model answers,
  116. with no tool call and no skill names to memorise. They are exclusive; running both would search
  117. twice for one turn.
  118. OpenClaw ships as two packages, because 2.0 dropped the hook the 1.x plugin injects through:
  119. `plugin-openclaw` for releases up to 2026.7.x, `plugin-openclaw2` for 2.0 (2026.8.1) and newer.
  120. The packaged Raven plugin installs and its on-demand mode works today; `mode: auto` there will
  121. claim the `skills` stage once Raven merges its upstream `context_segments` slot, and is inert
  122. until then. Raven's built-in retrieval keeps working either way.
  123. > Install SkillCorpus Plugins following https://github.com/EverMind-AI/SkillCorpus/blob/main/skillcorpus_plugin/INSTALL.agent.md
  124. Paste that line to your agent and it installs itself. Per-host setup, the five settings you
  125. will actually touch, what each turn costs and what leaves your machine —
  126. **[`skillcorpus_plugin/`](skillcorpus_plugin)**.
  127. <br>
  128. ## Public artifacts
  129. This is the concrete inventory of what is public today.
  130. | | Artifact | What | Link |
  131. |---|---|---|---|
  132. | 🌐 | **SkillHub** | the current 114,190-skill catalog + the two models, hosted as an API — no install | [evermind.ai/skillhub](https://evermind.ai/skillhub) |
  133. | 📚 | **Corpus** *(demo)* | the downloadable 1,000-skill sample — `skills.parquet` + `attachments.tar.zst` + dataset card; the full catalog is served by SkillHub | [🤗 demo-1k](https://huggingface.co/datasets/EverMind-AI/skillcorpus-demo-1k) |
  134. | 🗂️ | **Full corpus** *(coming soon)* | the complete 114,190-skill dataset, same schema and layout as the demo — `skills.parquet` + `attachments.tar.zst` + dataset card | 🤗 *coming soon* |
  135. | 🔡 | **Retrieval models** | a bi-encoder and a reranker, fine-tuned from `Qwen3-Embedding-0.6B` and `Qwen3-Reranker-0.6B` | [🤗 bi-encoder](https://huggingface.co/EverMind-AI/skillcorpus-embedding-0.6b) · [reranker](https://huggingface.co/EverMind-AI/skillcorpus-reranker-0.6b) |
  136. | 🛠️ | **Code** | this repo — the pipeline that builds the corpus and trains the two models (`aggregate` · `curate` · `match` · `evaluate` · `export`) | [GitHub](https://github.com/EverMind-AI/SkillCorpus) |
  137. | 🔌 | **Plugins** | packaged host adapters for OpenClaw · Hermes · WorkBuddy · Raven, plus DeepSeek Harness and an HTTP adapter | [`skillcorpus_plugin/`](skillcorpus_plugin) |
  138. *Open source today: the code, the 1,000-skill demo corpus, and the retrieval models. The full
  139. 114,190-skill corpus is being prepared for release on HuggingFace. The hosted
  140. SkillHub service itself stays closed.*
  141. <div align="center">
  142. <img src="https://github.com/user-attachments/assets/71edc5ab-291f-4fb8-8f5c-177d50e5b8f4" alt="16-class distribution over the 96,401 active skills" width="58%">
  143. </div>
  144. The 96,401-skill snapshot measured in the paper, organised by a 16-class taxonomy and three quality facets
  145. (utility / robustness / safety), with 1024-dim retrieval embeddings. Column contract:
  146. [`docs/corpus-schema.md`](docs/corpus-schema.md).
  147. <br>
  148. ## Query the API directly
  149. [SkillHub](https://evermind.ai/skillhub) serves the corpus in three tiers — discover
  150. (metadata), read (`skill_md`), download (zip with `scripts/`). Most skills are pure
  151. instructions, so the read tier is usually sufficient.
  152. ```bash
  153. curl "https://skillhub.evermind.ai/openapi/v1/skills?q=extract+tables+from+a+PDF"
  154. ```
  155. Take an `id` from the results, fetch its `skill_md`, and inject it into your agent's
  156. prompt. [`examples/skillhub_demo.py`](examples/skillhub_demo.py) runs all three tiers:
  157. ```bash
  158. # search + read the bodies — stdlib only, no install, no API key
  159. python examples/skillhub_demo.py "extract tables from a scanned PDF invoice"
  160. # also fetch the bundled scripts of the top hit
  161. python examples/skillhub_demo.py --install ./skills "convert a PDF to images"
  162. # retrieve AND run the task — any OpenAI-compatible LLM (OpenAI, OpenRouter, local vLLM, …)
  163. export OPENAI_API_KEY=... # OpenRouter / vLLM: also set
  164. # export OPENAI_BASE_URL=https://openrouter.ai/api/v1 # OPENAI_BASE_URL + --model openai/gpt-4o-mini
  165. python examples/skillhub_demo.py --ask "extract tables from a scanned PDF invoice"
  166. ```
  167. ```
  168. task: extract tables from a scanned PDF invoice
  169. [1/2] search → 2 hit(s), metadata only
  170. 1. ocr-and-documents q=0.808 DOC-PROC MIT
  171. Extract text from PDFs/scans (pymupdf, marker-pdf).
  172. 2. document-workflows q=0.86 DOC-PROC MIT
  173. Build end-to-end document processing workflows and pipelines …
  174. [2/2] detail → fetching skill_md for 2 skill(s)
  175. ocr-and-documents: 4916 chars u=8 r=7 s=9 files=4 flags=['no_steps']
  176. document-workflows: 31628 chars u=9 r=9 s=9 files=7
  177. → built a prompt of 36,742 chars with the skill bodies injected
  178. ```
  179. Endpoints, response envelope, status codes and rate limits:
  180. [`docs/integrations.md`](docs/integrations.md).
  181. <br>
  182. ## Self-host the models
  183. To avoid depending on the hosted endpoint, run selection yourself. The corpus and both
  184. retrieval models are released: load the data, serve the two models, and run your own
  185. encode → top-k → rerank.
  186. ```python
  187. # the data — a 1,000-skill demo for now; the full 114,190-skill corpus follows
  188. from datasets import load_dataset
  189. skills = load_dataset("EverMind-AI/skillcorpus-demo-1k", split="train") # 1,000 demo skills
  190. # or read the file directly with pandas (no `datasets`): pip install pandas
  191. import pandas as pd; skills = pd.read_parquet("skills.parquet")
  192. ```
  193. Attachments (`scripts/`, `references/`) ship as a sibling `attachments.tar.zst`.
  194. ```bash
  195. # install the serving deps (torch, transformers, …), then point the two env vars at
  196. # the released checkpoints (the script's defaults are training outputs absent from a
  197. # fresh clone) and serve both models behind one endpoint -> /embed + /score
  198. pip install -r skillcorpus/match/requirements.txt
  199. EMBEDDING_MODEL=<embedding checkpoint dir> RERANKER_MODEL=<reranker checkpoint dir> \
  200. bash skillcorpus/match/scripts/run_server.sh
  201. ```
  202. This endpoint speaks `/embed` + `/score`
  203. ([`skillcorpus/match/` → Serving](skillcorpus/match/README.md#serving)) — it is **not** a
  204. drop-in for SkillHub's hosted-only `/openapi/v1/skills` API. So:
  205. - `examples/skillhub_demo.py` and the section-C integrations talk only to the **hosted**
  206. SkillHub; a self-hosted setup runs its own selection directly over `/embed` + `/score`.
  207. - It is also the embedding endpoint the producer's dedup uses — set
  208. `embedding.provider: skillrouter_remote` to [build your own corpus](#build-your-own) with it.
  209. To curate **your own** sources instead, see [Build your own corpus](#build-your-own).
  210. <br>
  211. ## How it works
  212. <div align="center">
  213. <img src="https://github.com/user-attachments/assets/e0e72150-373b-4381-ad6c-74668d436d49" alt="SkillCorpus: curated skills are matched to a task and injected into an agent before execution" width="100%">
  214. <p><em>The collection pipeline is the foundation; the payoff is task-specific skill retrieval before the agent acts.</em></p>
  215. </div>
  216. ```
  217. skillcorpus/
  218. ├── core/ data models · SQLite/faiss store · LLM & embedding clients
  219. ├── aggregate/ source registry + multi-repo clone
  220. ├── curate/ parse · safety · license · classify · quality · dedup + full-library passes
  221. ├── export/ corpus writer (parquet + attachments + dataset card)
  222. ├── match/ the 2 released models + training recipe ← isolated deps
  223. ├── evaluate/ skillsbench · qwenclawbench · gdpval benchmarks ← isolated deps
  224. ├── lifecycle/ per-model skill retirement on the unaided level
  225. └── cli.py build · stats · export
  226. ```
  227. `cli build` runs the whole curation chain
  228. (`ingest → quality_pass → dedup_pass → license_audit → export.corpus`). LLM classification and
  229. quality scoring degrade gracefully to rules when no model endpoint is reachable, so the pipeline
  230. always runs end to end.
  231. `match/` and `evaluate/` are standalone toolkits with their own `requirements.txt`
  232. (torch / transformers, per benchmark); they are **not** pulled in by `pip install` of the producer.
  233. - **Retrieval** — [`skillcorpus/match/`](skillcorpus/match) is **the two released models**:
  234. a bi-encoder fine-tuned from `Qwen3-Embedding-0.6B` for candidate recall, and a reranker
  235. fine-tuned from `Qwen3-Reranker-0.6B` that scores the top candidates. SkillHub serves both;
  236. to run them yourself see [Serving](skillcorpus/match/README.md#serving) (`serve.py` +
  237. `run_server.sh`). The directory also holds the training
  238. recipe (synthetic queries → InfoNCE → listwise CE) and `eval_compare.py` for the retrieval
  239. metrics (nDCG / MRR / Hit / Recall).
  240. - **Benchmarks** — [`skillcorpus/evaluate/`](skillcorpus/evaluate): `skillsbench`,
  241. `qwenclawbench`, `gdpval` — each self-contained with its own README and dependencies.
  242. - **Retirement** — [`skillcorpus/lifecycle/`](skillcorpus/lifecycle) decides, per deployment
  243. model, which skills the model has internalized and can retire: it synthesizes a task from each
  244. skill, scores the model without and with the skill, and retires on the unaided level instead of
  245. the gain. It is the code of the paper *Beyond Gain: Measuring Skill Internalization for Agent
  246. Skill Retirement*.
  247. <a name="build-your-own"></a>
  248. <br>
  249. ## Build your own corpus
  250. Only needed if you want to curate **your own** sources. Requires an LLM endpoint for
  251. classification / quality scoring and an embedding endpoint for dedup — see
  252. [`docs/running.md`](docs/running.md).
  253. ```bash
  254. git clone https://github.com/EverMind-AI/SkillCorpus.git skillcorpus && cd skillcorpus
  255. python3 -m venv .venv && source .venv/bin/activate
  256. python -m pip install --upgrade pip && pip install -e .
  257. python -m skillcorpus.cli build # 4 demo sources -> curate -> export
  258. python -m skillcorpus.cli stats # counts by source / category / license
  259. python -m skillcorpus.cli export --out ./corpus
  260. ```
  261. Only skills from GREEN-licensed **sources** are exported (the demo trusts the whitelist in
  262. `audit/license_safe_sources.json` wholesale; production gates per source-repo SPDX). The per-row
  263. `license` is each skill's declared value, so a demo corpus can still carry non-GREEN `license`
  264. strings. Use `--sources-config your.yaml` for your own registry, or `--source <name>` for one source.
  265. ```bash
  266. pip install -e ".[dev]"
  267. python -m pytest skillcorpus/tests -p no:cacheprovider --import-mode=importlib
  268. ```
  269. <br>
  270. ## Roadmap
  271. <!-- TODO(@team): this is a first pass from known gaps — edit to match your plan. -->
  272. - [x] Curation pipeline: 16-class taxonomy, 3-facet quality, per-source license audit
  273. - [x] Fine-tuned retrieval stack + three-benchmark evaluation
  274. - [x] Per-model skill retirement on the unaided level (`skillcorpus/lifecycle/`)
  275. - [x] Public SkillHub endpoint
  276. - [x] Retrieval models (bi-encoder + reranker) and a 1k demo corpus on HuggingFace
  277. - [ ] Full 114,190-skill corpus on HuggingFace — coming soon
  278. - [x] Deployment script for the two retrieval models (self-hosting `match/`)
  279. - [x] Plugins for WorkBuddy · Hermes · OpenClaw · DeepSeek Harness (+ HTTP adapter for any other host)
  280. - [ ] Raven plugin — packaged, waiting on the upstream `context_segments` slot
  281. <br>
  282. ## EverMind Ecosystem
  283. EverMind connects memory research, production-ready products, and practical
  284. integrations into one open-source ecosystem.
  285. <table>
  286. <tr>
  287. <th colspan="2">Products</th>
  288. </tr>
  289. <tr>
  290. <td><strong><a href="https://github.com/EverMind-AI/EverOS">EverOS</a></strong></td>
  291. <td>A local-first, Markdown-native long-term memory runtime for agents and users.</td>
  292. </tr>
  293. <tr>
  294. <td><strong><a href="https://github.com/EverMind-AI/Raven">Raven</a></strong></td>
  295. <td>A memory-first, self-improving agent harness with proactivity, context control, and skill evolution.</td>
  296. </tr>
  297. <tr>
  298. <td><strong><a href="https://github.com/EverMind-AI/EverMe">EverMe (CLI)</a></strong></td>
  299. <td>A CLI and agent plugin suite for cross-device, cross-agent personal memory.</td>
  300. </tr>
  301. <tr>
  302. <th colspan="2">Research &amp; Evaluation</th>
  303. </tr>
  304. <tr>
  305. <td><strong><a href="https://github.com/EverMind-AI/SkillCorpus">SkillCorpus</a></strong></td>
  306. <td>Curated, retrieval-ready agent skill corpora with retrieval and evaluation tooling.</td>
  307. </tr>
  308. <tr>
  309. <td><strong><a href="https://github.com/EverMind-AI/EverAlgo">EverAlgo</a></strong></td>
  310. <td>Stateless extraction, ranking, parsing, and memory operators that power EverOS.</td>
  311. </tr>
  312. <tr>
  313. <td><strong><a href="https://github.com/EverMind-AI/HyperMem">HyperMem</a></strong></td>
  314. <td>Hypergraph-based hierarchical memory for coarse-to-fine long-term conversation retrieval.</td>
  315. </tr>
  316. <tr>
  317. <td><strong><a href="https://github.com/EverMind-AI/MSA">MSA</a></strong></td>
  318. <td>Memory Sparse Attention for scalable latent memory and 100M-token contexts.</td>
  319. </tr>
  320. <tr>
  321. <td><strong><a href="https://github.com/EverMind-AI/EverMemBench">EverMemBench</a></strong></td>
  322. <td>Evaluation of factual recall, applied reasoning, and personalized generalization in memory systems.</td>
  323. </tr>
  324. <tr>
  325. <td><strong><a href="https://github.com/EverMind-AI/EvoAgentBench">EvoAgentBench</a></strong></td>
  326. <td>Longitudinal evaluation of agent self-evolution, transfer efficiency, error avoidance, and skill use.</td>
  327. </tr>
  328. <tr>
  329. <th colspan="2"><a href="https://github.com/EverMind-AI/plugins">Integrations</a></th>
  330. </tr>
  331. <tr>
  332. <td><strong><a href="https://docs.openclaw.ai">OpenClaw</a></strong></td>
  333. <td><a href="https://github.com/EverMind-AI/plugins/tree/main/openclaw">OpenClaw plugin</a> for automatic recall, capture, and session-memory lifecycle management.</td>
  334. </tr>
  335. <tr>
  336. <td><strong><a href="https://github.com/NousResearch/hermes-agent">Hermes Agent</a></strong></td>
  337. <td><a href="https://github.com/EverMind-AI/plugins/tree/main/hermes">Hermes plugin</a> for persistent memory across Hermes sessions.</td>
  338. </tr>
  339. <tr>
  340. <td><strong><a href="https://github.com/deepseek-ai/DeepSeek-Harness">DeepSeek Harness</a></strong></td>
  341. <td><a href="https://github.com/EverMind-AI/plugins/tree/main/dsh">DSH plugin</a> for memory-aware DeepSeek Harness agents.</td>
  342. </tr>
  343. <tr>
  344. <td><strong><a href="https://dify.ai">Dify</a></strong></td>
  345. <td><a href="https://github.com/EverMind-AI/plugins/tree/main/dify">Self-hosted</a> and <a href="https://github.com/EverMind-AI/plugins/tree/main/dify_cloud">cloud</a> tools for explicit memory search and storage in workflows and agents.</td>
  346. </tr>
  347. </table>
  348. Together, these projects form EverMind's research-to-runtime stack: methods
  349. and benchmarks become reusable memory infrastructure, products, and agent
  350. integrations.
  351. ## Citation
  352. ```bibtex
  353. @article{wang2026skillcorpus,
  354. title = {SkillCorpus: Consolidating and Evaluating the Open Skill Ecosystem for Real-World LLM Agents},
  355. author = {Wang, Yanze and Yao, Pengfei and Sun, Tianyi and Hu, Chuanrui and Xiao, Yan and Luo, Xiaotian and Han, Yunyun and Chen, Yifan and Sun, Jun and Deng, Yafeng},
  356. year = {2026},
  357. eprint = {2607.15557},
  358. archivePrefix = {arXiv},
  359. url = {https://arxiv.org/abs/2607.15557}
  360. }
  361. ```
  362. <br>
  363. ## License
  364. - **Code** — Apache-2.0 (the `match/`, `evaluate/` and `lifecycle/` toolkits are each MIT — see their own `LICENSE`).
  365. - **Corpus** — every skill keeps its **original upstream license**; only GREEN
  366. (MIT / Apache-2.0 / BSD / ISC / …) skills are included, none relicensed. Each row carries
  367. `source`, `source_url`, and `license`, so downstream use must follow the per-skill terms.
  368. Full GREEN/RED/YELLOW policy, license data flow, and opt-out:
  369. [`docs/licence-and-governance.md`](docs/licence-and-governance.md).

Ratings and comments

Sign in to rate this skill.

Sign in

Free to use. If it’s helped, pass it on.