Skill

Appearance

SkillGive your AI agent superpowers through the terminal

A universal skill runtime that runs anywhere. No servers, no context bloat, just tools that work.

Get Started
View on GitHub
🚀

Single Binary

No servers required. Install once, works everywhere. Runs in CI/CD, terminal agents, or IDEs.

🔐

WASM Sandbox

Capability-based security. Skills can't escape their boundaries. Declare permissions upfront.

🔍

Local Semantic Search

Find tools using natural language. Zero context cost. Works offline. No API keys needed.

🌐

Universal Runtime

Same skills work with Claude Code, MCP protocol, or CLI commands. Write once, run everywhere.

The Problem ​

AI agents need to do things in the real world. Query databases. Deploy code. Manage infrastructure. Call APIs.

Today, you have two choices:

  1. Stuff tool docs into prompts → Context bloat, hallucinated flags, model confusion
  2. Run MCP servers → Complex setup, always-on processes, protocol overhead

Skill is a third way. A universal skill runtime that runs anywhere, stays secure, finds tools intelligently, and works with any agent.

Quick Start ​

bash
# Install
curl -fsSL https://dqkbk9o7ynwhxfjx.public.blob.vercel-storage.com/install.sh | sh

# Discover tools with natural language
skill find "manage kubernetes pods"

# Execute in sandboxed WASM
skill run kubernetes:get --resource pods --all-namespaces

Read the full installation guide →

Two Ways to Use ​

CLI Mode (Default) ​

Perfect for Claude Code, Aider, or any terminal-based agent:

bash
skill find "<what you want to do>"   # Discover tools
skill run <skill>:<tool> [args]      # Execute tools
skill list                           # See what's installed

The agent uses shell commands. No SDK, no protocol, no complexity.

MCP Server Mode ​

For agents that speak Model Context Protocol:

bash
# Start the MCP server
skill serve

Add to your Claude Desktop or IDE config:

json
{
  "mcpServers": {
    "skill": {
      "command": "skill",
      "args": ["serve"]
    }
  }
}

The same skills work in both modes. Write once, use everywhere.

Learn more about MCP integration →

Why Skill? ​

📝 Traditional Approach ​

Stuff docs into prompts

  • ❌ Context window bloat
  • ❌ Hallucinated command flags
  • ❌ Model confusion with complex tools
  • ❌ Slow, expensive API calls

🔌 MCP Servers ​

Always-on processes

  • ❌ Complex setup per tool
  • ❌ Memory overhead for servers
  • ❌ Protocol translation costs
  • ❌ Hard to debug failures

⚡ Skill Runtime ​

Best of both worlds

  • ✅ Zero context cost (local search)
  • ✅ Single binary, no servers
  • ✅ WASM sandbox security
  • ✅ <50ms tool discovery

Performance ​

MetricValue
Cold start~100ms (includes AOT compilation)
Warm start<10ms (cached)
Vector search<50ms (local FastEmbed)
MCP tool call<100ms typical

All operations run locally. No API calls. Works offline.

Security Model ​

Skills declare capabilities at install time and can't escape their boundaries:

yaml
capabilities:
  network:
    - "*.amazonaws.com"
  filesystem:
    - read: "${args.file}"
  allowed-tools:
    - kubectl
    - helm

What Skills Cannot Do ​

  • ❌ Read arbitrary files (WASI filesystem not mounted)
  • ❌ Access unrequested network (WASI sockets allowlist)
  • ❌ Run arbitrary commands (command allowlist)
  • ❌ Persist state (memory cleared after execution)

Who Uses Skill? ​

🤖 AI Agent Developers ​

Use Skill with Claude Code, Aider, Cursor, or custom agents. No integration code needed—just shell commands.

Quick start guide →

💻 DevOps Teams ​

Standardize tool access across your organization. Audit logging and RBAC built-in.

Enterprise features →

🛠️ Tool Developers ​

Write your skill once in Rust, JavaScript, or Python. Runs everywhere without modification.

Build a skill →

Get Started ​

Install Skill ​

Run your first skill in under 5 minutes.

Get Started →

Or explore:Build a Skill · API Reference · Browse Examples

Released under the Apache-2.0 License.

Copyright © 2024-present Skill Engine