ai-skills-sdk 0.2.0

Universal open standard for AI agent skills — write once, run anywhere

🧠 ai-skills

The universal open standard for AI agent skills — write once, run anywhere.

---

The Problem

AI agent frameworks are exploding — LangChain, AutoGen, CrewAI, Semantic Kernel, and dozens more. But skills (tools, functions, actions) built for one framework don't work in another. Every developer rewrites the same skills over and over, just in different formats.

This is the same problem the web solved with HTML. We need a common language for AI skills.

The Solution

ai-skills is an open specification + SDK that lets you:

1. Write a skill once in a simple skill.yaml file
2. Publish it to the public registry
3. Export it to any framework automatically

skill.yaml  →  LangChain tool
            →  AutoGen skill  
            →  CrewAI tool
            →  Semantic Kernel function
            →  Raw API call

How This Compares

Wondering how this compares to LangChain Tools, OpenAPI, MCP, or SKILL.md? Read our detailed breakdown: How ai-skills Compares.

---

Quick Start

Install the CLI

pip install ai-skills-sdk

Development install: If you want to contribute or run from source, clone the repo and install in editable mode:

git clone https://github.com/davyjones7321/AI-skills.git
cd AI-skills
pip install -e .

Create your first skill

aiskills init my-skill
cd my-skill

This creates a skill.yaml:

skill:
  id: my-skill
  version: 1.0.0
  name: My Skill
  description: What this skill does
  inputs:
    - name: input_text
      type: string
      required: true
  outputs:
    - name: result
      type: string
  execution:
    type: prompt
    prompt_template: "Do something with: {input_text}"

Validate and audit

# Check against the v0.1 schema specification
aiskills validate skill.yaml

# Run static security analysis (catch secrets & dangerous imports)
aiskills validate --audit skill.yaml

Export to your framework

aiskills export --target langchain    # → langchain_tool.py
aiskills export --target autogen      # → autogen_skill.py
aiskills export --target crewai       # → crewai_tool.py

Run a skill locally

# Dry-run: shows formatted prompt without calling any API
aiskills run skill.yaml --input '{"text": "hello"}'

# Run a code-type skill (no API key needed)
aiskills run examples/word-frequency/skill.yaml --input-file input.json

# Live execution: actually calls the LLM (requires OPENAI_API_KEY)
aiskills run skill.yaml --input-file input.json --execute

Publish to the registry

# Authenticate via GitHub OAuth (one-time)
aiskills login

# Preview what would be published (optional)
aiskills publish skill.yaml --dry-run

# Publish your skill to the registry
# Note: You can also publish via web upload or guided form at /publish
aiskills publish skill.yaml

Install a skill

# Optional: override the default hosted registry for this shell
export AISKILLS_REGISTRY_URL=https://ai-skills-sdk.onrender.com

# Download a published skill to your local workspace
aiskills install ai-skills-team/summarize-document

# Download and immediately auto-export it to LangChain!
aiskills install ai-skills-team/summarize-document --export langchain

Use the registry web app (MVP)

cd registry/frontend
npm install
npm run dev

Then open http://localhost:3000:

• / — marketing + discovery homepage with 12-card category grid
• /skills — browse/search with category filters and pagination
• /skills/{author}/{id} — full skill detail page
• /publish — interactive publishing studio (upload YAML or guided form)

---

Project Structure

ai-skills/
├── README.md               ← You are here
├── CONTRIBUTING.md         ← Contribution guidelines
├── CODE_OF_CONDUCT.md      ← Community standards
├── .github/
│   ├── workflows/ci.yml    ← CI: validate all skills on push/PR
│   ├── ISSUE_TEMPLATE/     ← Bug, feature, new skill templates
│   └── pull_request_template.md
├── docs/
│   ├── SPEC.md             ← The official v0.1 specification
│   ├── SECURITY.md         ← Security model and audit tools
│   └── COMPARISON.md       ← How we compare to MCP/LangChain/etc
├── examples/               ← 19 complete, working example skills (prompt, code, tool_call, chain)
│   ├── summarize-document/
│   ├── generate-sql/ 
│   ├── markdown-to-html/
│   └── ... (16 more)
├── sdk/
│   ├── cli.py              ← Main CLI entry point
│   ├── validator.py        ← Schema validation
│   ├── security.py         ← Security auditing
│   ├── runner.py           ← Skill execution engine (run command)
│   ├── auth_config.py      ← Authentication management
│   └── exporters/          ← Framework adapters
└── registry/
    ├── api/                ← FastAPI Registry Backend Server
    │   ├── categories.py   ← 12-category fixed taxonomy
    │   └── utils.py        ← Shared utils like make_json_safe
    ├── frontend/           ← Next.js registry web app (homepage, browse, detail, publish)
    │   ├── lib/skill-categories.ts  ← Shared category constants
    │   ├── lib/publish-utils.ts     ← Web publishing form utilities
    │   └── components/registry/publish-studio.tsx ← Interactive publish studio
    └── index.json          ← Prototype registry index

---

Skill Types

Type	Description	Example
prompt	LLM prompt template	Summarize, classify, translate
tool_call	External API/function call	Web search, database query
chain	Multiple steps in sequence	Research → summarize → format
code	Execute a code snippet	Data processing, calculations

---

Why ai-skills?

	ai-skills	LangChain	AutoGen	Semantic Kernel
Framework-agnostic	✅	❌	❌	❌
Open standard	✅	❌	❌	❌
Public registry	✅	❌	❌	❌
Built-in benchmarks	✅	❌	❌	❌
Skill portability	✅	❌	❌	❌

---

Contributing

We're in early days — contributions, feedback, and ideas are very welcome.

• 📖 Read the Specification
• 🛠️ Check out example skills
• 💬 Open an issue to discuss ideas
• 🔁 Submit a PR

---

License

MIT — free to use, modify, and distribute.

---

Session Update — 2026-03-26

GitHub OAuth implementation for the registry was completed in this session across the backend, frontend, and CLI. The backend now uses signed JWTs with proper verification, GitHub OAuth state validation, frontend callback redirects, stateless logout, and user email plus last-login persistence. The frontend now includes /login, /auth/callback, an auth context/provider, and dynamic signed-in header state. The CLI aiskills login flow was updated to complete browser-based auth without manual token paste, and registry/api/.env.example was added with the required auth variables.

Session Update — 2026-03-27

The hosted registry configuration was tightened up in this session. The backend GitHub OAuth callback URL is now built from BASE_URL, so production callbacks point back to the backend itself instead of the frontend. The API docs and example env file were updated with BASE_URL, the backend CORS allowlist now includes https://ai-skills-omega.vercel.app, and the CLI registry commands now honor AISKILLS_REGISTRY_URL with a default fallback of https://ai-skills-sdk.onrender.com.

Session Update — 2026-03-27 (Production Hardening)

A comprehensive production audit was run and 10 issues were fixed: insecure default secrets removed (now required fields), debug defaults to False, CORS origins simplified to settings.frontend_url, OAuth state moved from in-memory dict to a persistent DB table (OAuthState), getattr removed from the code sandbox, deprecated @app.on_event("startup") replaced with modern lifespan, health check no longer exposes environment in production, duplicate DEFAULT_REGISTRY_URL import consolidated, and .env.example updated with all production variables.

Session Update — 2026-03-28 (Standalone CLI)

The SDK is now published to PyPI as ai-skills-sdk. Users can install it directly with pip install ai-skills-sdk — no repo clone required. All documentation across README.md, OVERVIEW.md, CONTRIBUTING.md, registry/README.md, registry/api/README.md, and docs/LAUNCH_POST.md has been updated to reflect this. The frontend publish page now shows the correct install steps and includes a clone notice for development contributors. The auth context bug (header not updating after OAuth callback) was also fixed.