Agent Skills SDK

Implementation of Anthropic's Agent Skills open standard - makes skills discoverable and usable by agents in production frameworks.

What is Agent Skills?

Agent Skills is an open standard released by Anthropic for packaging agent capabilities in a composable, portable format.

Skills = Organized folders with instructions, scripts, and resources that agents can discover and load dynamically.

The Format:

• SKILL.md file with YAML frontmatter (metadata)
• Progressive disclosure: Level 1 (metadata) → Level 2 (instructions) → Level 3+ (references)
• Enables agents to specialize without rebuilding entire systems

Already in use by Anthropic and companies in production.

What This SDK Does

This SDK implements the Agent Skills standard, providing the tooling layer to make skills work in real-world applications:

• ✅ Discovery - Find and parse skills in Agent Skills format
• ✅ Dynamic Loading - Progressive disclosure (metadata first, full content on-demand)
• ✅ Framework Integration - Use with Agno, LangChain, CrewAI, custom frameworks
• ✅ Token Optimization - Load only what agents need (85-95% savings)
• ✅ Production Ready - Pip installable, tested, ready to deploy

Quick Start

# Install
pip install -e .

# Set API key (for examples)
export OPENROUTER_API_KEY="your-key"

# Run examples
python examples/python/agent.py      # Pure Python (no framework)
python examples/agno/agent.py        # Agno framework
python examples/validate.py          # Token savings validation

Basic Usage

from agno.agent import Agent as Agno
from agno.models.openrouter import OpenRouter
from agent_skills_sdk.adapters import AgnoAdapter

# Create adapter - discovers skills in Agent Skills format
adapter = AgnoAdapter(skill_paths=["./skills"])

# Create agent
agent = Agno(
    name="assistant",
    model=OpenRouter(id="anthropic/claude-3.5-sonnet")
)

# Attach skills - agent can now discover and load them dynamically
adapter.attach_to_agent(agent)

# Use it
response = agent.run("Help me extract text from a PDF")
# Agent discovers PDF skill and loads it on-demand

How It Works

The SDK provides 4 tools to agents for intelligent skill management:

1. list_available_skills() - See all available skills (metadata only, ~1600 tokens)
2. search_skills(query) - Find relevant skills by keyword
3. load_skill_instructions(name) - Load specific skill on-demand (~2000 tokens)
4. get_skill_tools(name) - Get tool information

Progressive Disclosure in Action:

User: "Can you extract text from a PDF?"
  ↓
Agent: Calls list_available_skills()
  → Sees 16 skills (metadata only: ~1600 tokens)
  ↓
Agent: Calls search_skills("pdf")
  → Finds "pdf" skill
  ↓
Agent: Calls load_skill_instructions("pdf")
  → Loads PDF instructions (~2000 tokens)
  ↓
Agent: Uses PDF skill to help user

Total: ~3600 tokens
vs Loading all 16 skills upfront: ~36,000 tokens
Savings: 90%

This implements the progressive disclosure pattern from the Agent Skills specification.

Skills Included

16 skills following the Agent Skills format:

Office Suite

• pdf - PDF manipulation and text extraction
• xlsx - Excel spreadsheet creation and analysis
• pptx - PowerPoint presentation creation
• docx - Word document processing

Creative

• canvas-design - Visual design and graphics
• algorithmic-art - Generative art creation
• theme-factory - Theme and style generation

Development

• frontend-design - Frontend development and UI
• webapp-testing - Web application testing
• web-artifacts-builder - Build web artifacts

Communication

• slack-gif-creator - Create GIFs for Slack
• internal-comms - Internal communications
• doc-coauthoring - Collaborative document editing

Meta

• skill-creator - Create new Agent Skills
• mcp-builder - Build MCP servers
• brand-guidelines - Brand guideline management

All skills follow the Agent Skills specification.

Examples

The SDK includes examples for multiple frameworks:

Agno Framework

python examples/agno/agent.py

Agno agent with dynamic Agent Skills loading. Requires OPENROUTER_API_KEY.

LangChain

python examples/langchain/agent.py

LangChain agent with Agent Skills integration. Requires OPENAI_API_KEY and pip install langchain langchain-openai.

CrewAI

python examples/crewai/agent.py

CrewAI agent with Agent Skills. Requires OPENAI_API_KEY and pip install crewai.

Pure Python (No Framework)

python examples/python/agent.py

Direct SDK usage without any agent framework. No API key required - demonstrates the core SDK functionality.

Validate Token Savings

python examples/validate.py

See the actual token savings from progressive disclosure (84-91% reduction).

Token Savings

Measured with real Agent Skills:

Scenario	Tokens Used	vs All Skills	Savings
All 16 skills loaded upfront	36,884	-	0%
Dynamic loading (1 skill)	3,282	36,884	91%
Dynamic loading (2 skills)	5,805	36,884	84%

Progressive disclosure enables massive token savings while maintaining full capability.

Framework Adapters

Agno

from agent_skills_sdk.adapters import AgnoAdapter

adapter = AgnoAdapter(skill_paths=["./skills"])
adapter.attach_to_agent(agent)

LangChain

from agent_skills_sdk.adapters import LangChainAdapter

adapter = LangChainAdapter(skill_paths=["./skills"])
tools = adapter.get_tools()

CrewAI

from agent_skills_sdk.adapters import CrewAIAdapter

adapter = CrewAIAdapter(skill_paths=["./skills"])
tools = adapter.get_tools()

Creating Skills

Skills follow the Agent Skills format:

---
name: my-skill
description: What this skill does
category: utility
---

# Skill: My Skill

## Overview
Instructions for the AI agent...

## Tools Available
- tool_name: Description of the tool

## Examples
...

Place in: skills/my-skill/SKILL.md

See the Agent Skills specification for complete format details.

Validation

Verify token savings with your skills:

python examples/validate.py

Example output:

✅ Discovered 16 skills
✅ Token savings: 84.3% (31,079 tokens saved)
✅ All skills loaded: ~36,884 tokens
✅ Dynamic loading: ~5,805 tokens

API Reference

AgnoAdapter

from agent_skills_sdk.adapters import AgnoAdapter

adapter = AgnoAdapter(skill_paths: List[str])
adapter.attach_to_agent(agent)
adapter.get_token_usage_stats()  # Get usage statistics

AgentSkillsClient

from agent_skills_sdk import AgentSkillsClient

client = AgentSkillsClient(
    skill_paths: List[str],
    auto_discover: bool = True
)

# Discovery
skills = client.discover_metadata()  # Lightweight metadata only

# Loading
skill = client.load_skill(name: str)  # Load full skill
instructions = client.get_instructions(name: str)  # Instructions only

# Search
results = client.search_skills(query: str)  # Find relevant skills

Project Structure

agentskills/
├── src/agent_skills_sdk/    # SDK implementation
│   ├── client.py             # Skills client
│   ├── discovery.py          # Skill discovery
│   ├── parser.py             # Agent Skills format parser
│   └── adapters/             # Framework adapters
│       ├── agno.py           # Agno adapter
│       ├── langchain.py      # LangChain adapter
│       └── crewai.py         # CrewAI adapter
├── skills/                   # 16 example skills (Agent Skills format)
├── examples/                 # Framework examples
│   ├── agno/agent.py         # Agno framework
│   ├── langchain/agent.py    # LangChain framework
│   ├── crewai/agent.py       # CrewAI framework
│   ├── python/agent.py       # Pure Python (no framework)
│   └── validate.py           # Token savings validation
├── tests/                    # Unit tests
└── README.md

Development

# Install dependencies
pip install -e .

# Run tests
pytest tests/

# Run validation
python examples/validate.py

Relationship to Agent Skills

This SDK implements the Agent Skills specification:

• Agent Skills (by Anthropic) = The open standard/format
• Agent Skills SDK (this project) = Production implementation with framework integrations

Like:

• Markdown = Format specification
• Markdown parsers = Implementations

We follow the Agent Skills specification and will evolve as the standard evolves.

Related Resources

• Agent Skills Announcement - Anthropic's introduction
• Agent Skills Specification - Official specification
• Agent Skills Repository - Reference implementation
• Example Skills - Community skills

License

MIT License - See LICENSE file

Credits

Agent Skills Standard: Created by Anthropic as an open standard.

SDK Implementation:

• Author: Supreeth Ravi
• Website: supreethravi.com
• LinkedIn: linkedin.com/in/supreeth-ravi
• Twitter/X: @supreeth___ravi
• Email: @supreethravi

This SDK provides implementation and tooling for the Agent Skills ecosystem.

Support

• Issues: GitHub Issues
• Documentation: This README
• Examples: See examples/ directory