A Living Knowledge Infrastructure for Agent Skills
Just as Wikipedia organizes human knowledge and GitHub governs software, SkillWiki provides the missing infrastructure for agent skills. It transforms heterogeneous knowledge sources — trajectories, documents, API specs, and execution experience — into reusable skill assets that are grounded in their originating evidence, version-controlled, and continuously evolved. SkillWiki covers the complete skill lifecycle: from knowledge ingestion and skill production, through provenance-aware graph exploration and human governance, to execution-driven health monitoring and automated evolution. The result is a shared infrastructure where knowledge, skills, and execution experience co-evolve together.
📺 Demo Video (English) · 📺 Demo Video (Chinese) · 🚀 Live System (HF Space) · 📖 Paper
SkillWiki system overview — from multi-source ingestion to governed skill evolution.
| Feature | Description | |
|---|---|---|
| 📥 | Multi-source ingestion | Trajectories, documents, API specs, scripts, and existing skill JSON/JSONL files |
| 🔄 | Governed lifecycle | Eight-stage state machine (S0–S7) with automated verification and full audit trail |
| 🕸️ | Provenance knowledge graph | Typed edges — depends_on, composes_with, evolved_from, replaces, and more |
| 🩺 | Health & evolution engine | Continuous monitoring detects degraded skills and queues repair proposals automatically |
| 🌐 | Bilingual UI | English (default) and Chinese — toggle via the language button in the header |
| ⌨️ | CLI-first | Every operation available as a skillwiki command for scripting and agent integration |
S0 Raw ──► S1 Candidate ──► S2 Draft ──► S3 Verified ──► S4 Released
│
S7 Archived ◄── S6 Deprecated ◄── S5 Degraded
| State | Name | Meaning |
|---|---|---|
| S0 | Raw | Ingested source material, not yet extracted |
| S1 | Candidate | Extracted skill candidate awaiting review |
| S2 | Draft | Formalized schema, pending verification |
| S3 | Verified | Passed automated postcondition checks |
| S4 | Released | Approved for agent use in production |
| S5 | Degraded | Success rate fell below health threshold |
| S6 | Deprecated | Replaced or retired |
| S7 | Archived | Read-only historical record |
- Python 3.10+
- Node.js 18+ (frontend only)
git clone https://github.com/Huangdingcheng/SkillWiki.git
cd SkillWiki/skillwiki
python -m venv venv
source venv/bin/activate # Linux / Mac
venv\Scripts\activate # Windows
pip install -r requirements.txt
pip install -e .skillwiki serve --port 8001 --api-key YOUR_LLM_API_KEYcd ../skillwiki-frontend
npm install
npm run dev # → http://localhost:3000Double-click START_SKILLWIKI_DEMO.bat in the repository root. On first run you will be prompted for your LLM endpoint and API key (saved to skillwiki-launcher\config.local.ps1, which is Git-ignored — never commit it).
Backend: http://127.0.0.1:8001
Frontend: http://127.0.0.1:5174/wiki
Restore the public demo fixtures after a restart:
RESTORE_SKILLWIKI_DEMO_STATE.bat
| Step | Page | What to explore |
|---|---|---|
| 1 | /wiki |
Skill library — browse, filter by state / tag, inspect schemas |
| 2 | /ingest |
Paste an API doc or upload a file; watch candidates extracted live |
| 3 | /graph |
Provenance knowledge graph — Nebula, Readable, and Debug view presets |
| 4 | /harness |
Execute-verify loop; observe automated repair attempts |
| 5 | /evaluation |
SkillsBench P0 sparse-subset analysis |
| 6 | /versions |
Business-readable diffs and re-verification after changes |
Install once after cloning:
cd skillwiki && pip install -e .All commands accept --api-url <URL> (default http://127.0.0.1:8001).
skillwiki serve [--host HOST] [--port PORT] [--backend memory|sqlite|postgres]# source_type: trajectory | document | api_doc | script | past_skills
skillwiki ingest run <source_type> <input> [--create]
skillwiki ingest run api_doc ./openai_spec.md --create
skillwiki ingest run trajectory "open browser -> search -> copy link"
skillwiki ingest run past_skills ./skills.json --max-candidates 20
skillwiki ingest status <candidate_id>skillwiki skill list [--state S3] [--tag nlp] [--limit 20]
skillwiki skill get <skill_id> [--full]
skillwiki skill status <skill_id>
skillwiki skill exec <skill_id> --input '{"key": "value"}'
skillwiki audit <skill_id>
skillwiki verify <skill_id> [--harness mock|claude_code|codex] [--max-retries 3] [--watch]
skillwiki promote <skill_id> <target_state>skillwiki health [--json] # system-wide overview
skillwiki health <skill_id> [--json] # success rate, issues, open proposals
skillwiki repair <skill_id> # generate maintenance candidate
skillwiki evolve [--json] # run one full evolution cycleskillwiki proposal list [--status pending|accepted|rejected] [--json]
skillwiki proposal accept <proposal_id>
skillwiki proposal reject <proposal_id>skillwiki graph neighbors <skill_id> [--depth 1]
skillwiki graph show <skill_id> [--view skill_only|provenance|version_impact] [--depth 2]
skillwiki graph deps <skill_id>
skillwiki graph export <skill_id> [-o output.json] [--view provenance] [--depth 2]skillwiki run "summarize the attached PDF and extract action items" [--verbose]- Knowledge-to-skill transformation from heterogeneous sources.
- Evidence-grounded skill assets with explicit provenance.
- Multi-granularity skill organization and dependency management.
- Provenance-aware exploration and visualization.
- Versioned skill governance with interpretable evolution.
- Execution-driven verification and continuous improvement.
- End-to-end demonstrations of reusable skill ecosystems.
| Variable | Default | Description |
|---|---|---|
LLM_API_KEY |
— | LLM API key (required) |
LLM_API_URL |
https://api.deepseek.com |
LLM base URL |
LLM_MODEL |
deepseek-v4-flash |
Model name |
SKILLOS_API_TARGET |
http://127.0.0.1:8001 |
Frontend proxy target |
MIT © 2026 SkillWiki Authors