SkillCapture 🧠

A privacy-first, local AI agent that watches your daily chats, automatically learns your repetitive workflows, and turns them into one-click Skills — all stored safely on your own hard drive.

Built with FastMCP · Works with Claude Desktop, Cursor, Windsurf, and any MCP-compatible client.

How It Works

SkillCapture uses a two-tier pipeline inspired by how human memory consolidation works:

Day 1 — Lightweight Draft (Cheap)

The AI scans your chat log and extracts potential workflows into a flat JSON cache. No heavy processing — just keywords and action summaries.

Day 2 — Heavy Promotion (Only on Match)

If you repeat a workflow, the system detects the keyword overlap and only then triggers the expensive generation: building a full, reusable Skill with named variables, step-by-step actions, and trigger phrases.

DISCOVERED → PENDING → PROMOTED → DEPRECATED
   (Day 1)    (Cache)   (Vault)    (30d unused)

The Storage Architecture

Skills are stored as Markdown files — you can read, edit, and version-control them with Git.

Quick Start

1. Install

Python (uvx / pipx) 🐍

uvx skill-capture-mcp
# or
pipx install skill-capture

2. Configure your LLM provider

cp .env.example .env
# Edit .env with your provider and API key

SkillCapture ships with three built-in providers. Set LLM_PROVIDER in .env:

Extensible: Need a different provider? Implement the LLMClient.chat() interface in core/providers.py.

3. Run the MCP Server

skill-capture-mcp
# (or `npx @YOUR_USERNAME/skill-capture-mcp`)

Then connect from Claude Desktop, Cursor, Windsurf, or any MCP-compatible client.

4. Connect to your MCP client

{
  "mcpServers": {
    "skill-capture": {
      "command": "python",
      "args": ["/absolute/path/to/skill-capture/server.py"],
      "env": { "LLM_PROVIDER": "openai", "OPENAI_API_KEY": "sk-..." }
    }
  }
}

{
  "mcpServers": {
    "skill-capture": {
      "command": "skill-capture-mcp",
      "args": [],
      "env": { "LLM_PROVIDER": "openai", "OPENAI_API_KEY": "sk-..." }
    }
  }
}

{
  "mcpServers": {
    "skill-capture": {
      "command": "skill-capture-mcp",
      "args": [],
      "env": { "LLM_PROVIDER": "openai", "OPENAI_API_KEY": "sk-..." }
    }
  }
}

codex mcp add skill-capture -- skill-capture-mcp

[mcp_servers.skill-capture]
type = "stdio"
command = "skill-capture-mcp"
args = []

[mcp_servers.skill-capture.env]
LLM_PROVIDER = "openai"
OPENAI_API_KEY = "sk-..."

CLI Mode

Don't need MCP? Use SkillCapture standalone from the terminal:

skill-capture-cli analyze           # Run the Day 1/Day 2 pipeline
skill-capture-cli list              # List all promoted skills
skill-capture-cli pending           # View pending drafts in the sandbox
skill-capture-cli run "Deploy App"  # Load and display a specific skill

MCP Tools

Once connected, your AI client has access to these tools:

Project Structure

skill-capture/
├── data/
│   └── pending.json          # The Sandbox
├── skills/
│   ├── index.json            # The Index
│   └── *.md                  # The Vault
├── logs/                     # Daily chat logs (input)
├── core/
│   ├── models.py             # Two-tier Pydantic schemas
│   ├── storage.py            # File-system I/O layer
│   ├── evaluator.py          # LLM client interface + evaluator logic
│   ├── providers.py          # OpenAI, Anthropic, Gemini clients
│   └── scheduler.py          # APScheduler nightly worker
├── server.py                 # FastMCP server
├── cli.py                    # Standalone CLI interface
└── requirements.txt

Tech Stack

• Python — Core language
• FastMCP — Model Context Protocol server framework
• Pydantic — Structured data validation
• OpenAI · Anthropic · Google Gemini — LLM providers (swappable)
• python-frontmatter — Markdown + YAML parsing
• APScheduler — Background task scheduling

Contributing

Contributions are welcome! Some ideas:

• 🔌 Add more LLM providers (Ollama, local models)
• 🎨 Build the web UI for skill management
• 📊 Add usage analytics and skill effectiveness tracking
• 🧪 Improve the keyword matching with embeddings
• 📝 Add support for more chat log formats

License

MIT License — see LICENSE for details.