zjunlp/skillnet

---

SkillNet provides unified infrastructure for the agent skill lifecycle:

• Discovery: search a public skill library by keyword or semantic intent.
• Installation: download skill folders from GitHub into local agent workspaces.
• Creation: generate structured skills from repositories, documents, prompts, or execution traces.
• Evaluation: assess skills for safety, completeness, executability, maintainability, and cost awareness.
• Analysis: extract capabilities and usage scenarios, and infer relationships between local skills.
• Routing: select skills for a task from your local library, with selection reasons.

---

News

• [2026-09-24] skillnet-ai 0.1.2 adds a local browser interface for exploring skill folders, analyzed relationships, and source evidence. Launch it with skillnet ui.

• [2026-08-20] The updated SkillNet report presents SkillNet-Gym, with executable benchmarks for skill construction, retrieval, and composition, and SkillNet-Fabric, which routes tasks through a Wiki built for each task.

• [2026-07-11] The SkillNet library now indexes 500K+ GitHub skills, with improved deduplication and broader coverage of scientific research and data analysis. This update also adds local scenario graphs and orchestration.

• [2026-03-26] JiuwenClaw integrates SkillNet as a built-in skill marketplace, bringing skill discovery and installation into its agent workflow.

• [2026-03-12] The SkillNet MCP server, maintained by CycleChain, makes SkillNet tools available to MCP-compatible agents.

• [2026-03-04] SkillNet: Create, Evaluate, and Connect AI Skills is now available on arXiv, describing the project’s approach to reusable agent skills.

• [2026-02-23] OpenClaw now includes SkillNet as a built-in skill for discovering and reusing agent capabilities.

---

Web Platform

Explore the public skill library on skillnet.openkg.cn.

• Find skills: search the skill library by keyword or semantic intent, and filter by category.
• Review skills: inspect descriptions, evaluation ratings, and GitHub sources, with download links for reuse.
• Explore collections: browse curated skill collections for specific domains and tasks, and inspect the relationships between their skills.

The site also introduces SkillNet-Gym for skill graph construction and lifecycle benchmarking, and SkillNet-Fabric for Wiki-based skill routing, including research results and a guided routing demo.

---

Local Browser Interface

The browser interface is part of the Python package. Browse your own skill folders, inspect saved analysis graphs, and follow relationship evidence to the original source lines.

Available in skillnet-ai 0.1.2+. Install and launch it with:

pip install "skillnet-ai[ui]>=0.1.2"
skillnet ui --skills-dir "/absolute/path/to/skills"

Open http://127.0.0.1:8765. Choose Current skill files to read the sources or Analysis results to explore saved compose_with and similar_to relationships. You can also import a graph.json generated by skillnet analyze.

Packaged releases include the website; users do not need Node.js or model keys to browse. Generate analysis separately with the SDK or CLI, then view the same results here. The interface starts with an empty personal library.

---

Python SDK

Install

Requires Python 3.10 or newer.

pip install skillnet-ai

Install optional dependencies for analysis and routing:

pip install "skillnet-ai[graph]"         # scenario analysis
pip install "skillnet-ai[graph,claude]"  # analysis and routing via Claude
pip install "skillnet-ai[graph,codex]"   # analysis and routing via Codex

Initialize

from skillnet_ai import SkillNetClient

client = SkillNetClient()

The client reads environment variables and saved settings. See Configuration to set up model endpoints or GitHub authentication.

Search

Search the hosted SkillNet catalog by keyword or semantic intent. Keyword search defaults to sorting by stars; vector search uses the search service’s embeddings.

results = client.search(
    q="analyze financial PDF reports",
    mode="vector",
    threshold=0.85,
    limit=10,
)

for skill in results:
    print(skill.skill_name, skill.stars, skill.skill_url)

Download

Download a GitHub skill folder and its resources, check its structure, and return the installed path. Use overwrite=True to replace an existing folder.

local_path = client.download(
    url="https://github.com/anthropics/skills/tree/main/skills/pdf",
    target_dir="./my_skills",
)
print(local_path)

Create

Generate skill packages from a prompt, repository, document, or execution trace. Choose one source per call; the result is a list of generated directory paths.

paths = client.create(
    prompt=(
        "Create a csv-quality-checker skill that checks CSV files for missing "
        "values and duplicate rows without modifying the input."
    ),
    output_dir="./my_skills",
)
print(paths)

client.create(
    github_url="https://github.com/zjunlp/DeepKE",
    output_dir="./my_skills",
)

client.create(
    office_file="./guide.pdf",
    output_dir="./my_skills",
)

Evaluate

Assess a local skill or GitHub skill URL across five quality dimensions. Each dimension contains a level (Good, Average, or Poor) and a reason. Evaluation reviews the skill’s instructions and supporting files without executing its scripts.

report = client.evaluate("./my_skills/pdf")
print(report["safety"]["level"], report["safety"]["reason"])
print(report["maintainability"]["level"], report["maintainability"]["reason"])

Analyze

Extract scenarios and capabilities from local skills, then build a reusable relationship graph, retrieval index and Wiki for routing.

analysis = client.analyze("./my_skills", output_dir="./skillnet_index")
print(analysis.index_dir)
print(analysis.skill_count, analysis.relation_counts)

The graph has two relations: directed compose_with for an output or state from one skill that supports another, and undirected similar_to for comparable capabilities. Both are tied to specific scenarios.

analyze reads direct child folders containing SKILL.md. Its default output is /.skillnet; pass output_dir to choose another location. Reanalyze after editing skills to update the snapshot used by routing. See analysis options and index details.

Route

Select up to k skills for a task using an index created by analyze. Hybrid search and graph expansion identify candidates; a Claude or Codex Agent SDK compares them in a task Wiki containing skill profiles, relationships, and source text.

result = client.route(
    "Extract tables from my PDF report and check the resulting CSV "
    "for missing values and duplicate rows.",
    index_dir="./skillnet_index",
    k=5,
)
for skill in result.skills:
    print(skill.skill_id, skill.name, skill.path, skill.reason)

Relations guide candidate exploration; the task and each skill’s constraints determine the final selection.

route returns skills (skill_id, name, path, reason) and available SDK usage. It selects up to k skills, possibly none. The selected set may cover only part of a task, and its order does not specify execution order.

See routing options and endpoint configuration.

CLI

The CLI exposes the same six operations. Use python -m skillnet_ai wherever the skillnet command is unavailable on your PATH.

Use skillnet --help for full options, or see the CLI examples. Add --json for a structured {ok, data, error} response on stdout; logs go to stderr.

create checks the generated package structure; add --evaluate to request a model assessment as well. Use skillnet validate for a local structure check without a model call.

Configuration

Search and public downloads need no API key. Create, evaluate, and analyze use an OpenAI-compatible Chat Completions endpoint. Analyze also needs an embedding endpoint; route uses that same embedding endpoint and a separately configured Claude or Codex Agent SDK.

Settings resolve in this order: explicit arguments → environment variables → user configuration → defaults. Load any .env file into your environment before calling the SDK; it does not load one automatically.

Linux and macOS:

export API_KEY="your-api-key"
export BASE_URL="https://api.openai.com/v1"
export SKILLNET_MODEL="gpt-4o"

Windows PowerShell:

$env:API_KEY = "your-api-key"
$env:BASE_URL = "https://api.openai.com/v1"
$env:SKILLNET_MODEL = "gpt-4o"

Analysis and routing require additional settings from the table above. See endpoint configuration for complete examples. skillnet configure --interactive saves settings locally; skillnet doctor --json checks configuration and optional dependencies.

---

REST API

The SkillNet search API is public and requires no authentication.

curl "http://api-skillnet.openkg.cn/v1/search?q=pdf&sort_by=stars&limit=5"
curl "http://api-skillnet.openkg.cn/v1/search?q=reading%20charts&mode=vector&threshold=0.8"

---

Use SkillNet Inside Agents

Install the SkillNet skill to give your agent access to all six operations. See installation and configuration and agent setup for instructions.

The demo below shows Claude Code using the SkillNet skill.

Model Context Protocol

The community SkillNet MCP server, maintained by CycleChain, wraps the SkillNet CLI. It requires Python, Node.js, and an installed skillnet-ai package.

git clone https://github.com/CycleChain/skillnet-mcp
cd skillnet-mcp
npm install

Docker:

docker pull fmdogancan/skillnet-mcp:latest

Follow the MCP setup guide to register the server with your agent and check its available tools and configuration requirements.

OpenClaw and JiuwenClaw

OpenClaw includes SkillNet as a built-in skill; JiuwenClaw integrates it into its skill marketplace. See the JiuwenClaw guide.

The demo below shows SkillNet running inside OpenClaw to discover and use reusable skills.

---

Examples and Experiments

Scientific discovery

The scientific workflow notebook illustrates skill discovery and reuse for scRNA-seq analysis, pathway lookup, and target validation. It uses a predefined plan and simulated data for parts of the demonstration.

Open the scientific workflow notebook.

More examples and benchmarks

• examples/: SDK demos and notebook workflows.
• experiments/: reproduction scripts for ALFWorld, WebShop, and ScienceWorld.

Complete the benchmark environment setup before running the experiments. Each benchmark has separate environment and data dependencies.

cd experiments

python alfworld_run.py --model o4-mini --split dev --max_workers 10 --exp_name alf_test --use_skill
python scienceworld_run.py --model o4-mini --split test --max_workers 5 --exp_name sci_test --use_skill
python webshop_run.py --model o4-mini --max_workers 3 --exp_name web_test --use_skill

---

Roadmap

• Broader evaluation of task routing across local skill libraries.
• More curated skill collections and routing Wikis.
• Stronger skill evaluation and regression testing.
• Extend SkillNet-Fabric routing across skill collections.
• Expand SkillNet-Gym lifecycle evaluation and training environments.

---

Contributing

Contributions are welcome: bug fixes, documentation, examples, integrations, and new skills all help. Please keep pull requests focused and include reproduction steps or examples when possible.

See the SDK development guide for code organization and local checks.

---

Citation

If SkillNet is useful in your research or agent system, please cite:

@article{liang2026skillnet,
  title={{SkillNet}: Create, Evaluate, and Connect {AI} Skills},
  author={Liang, Yuan and Zhong, Ruobin and Xu, Haoming and Jiang, Chen and Zhong, Yi and Fang, Runnan and Gu, Jia-Chen and Deng, Shumin and Yao, Yunzhi and Wang, Mengru and others},
  journal={arXiv preprint arXiv:2603.04448},
  year={2026}
}

---

License

SkillNet is licensed under MIT. Skills from external repositories retain their own licenses.