AutoSkill

Local knowledge graph + skill deduplication for AI agents
_{Drop files in, get organized knowledge out — zero cloud dependency}

---

AutoSkill watches a folder for AI agent skills and documents, deduplicates overlapping knowledge, builds a local vector-indexed knowledge graph, and proposes new skills — all running on your machine with a local LLM.

Features

• Skill Deduplication — Detects overlapping skills via embeddings, extracts shared knowledge with LLM
• Knowledge Indexing — PDF, Markdown, DOCX, HTML, TXT parsed, chunked, and stored in ChromaDB
• Auto-Tagging — Two-tier: high-similarity inherits neighbor tags, otherwise local LLM assigns them
• Skill Proposals — Conservative LLM evaluation proposes new Claude Code skills from your documents
• Live Dashboard — Real-time skill graph, 3D embedding space, pipeline status
• Obsidian Integration — Output folder is your vault; bundled plugin shows tags, proposals, and search
• MCP Server — Claude Code queries your knowledge graph for passive augmentation
• Fully Local — Embeddings via sentence-transformers, LLM via oMLX / any OpenAI-compatible API

Quick Start

uv sync
cp .env.example .env

# Index documents
uv run autoskill index input_skills/*.md input_skills/*.pdf --vault input_skills --output output

# Deduplicate skills
uv run autoskill ingest input_skills --output output

# Launch dashboard
uv run autoskill dashboard
# → http://localhost:8420

Dashboard

Interactive skill graph with shared knowledge nodes, 3D embedding scatter plot, and live pipeline progress.

How It Works

flowchart LR
    subgraph Input["input_skills/"]
        S["SKILL.md folders"]
        D["PDF / MD / DOCX / TXT"]
    end

    subgraph Pipeline["AutoSkill"]
        direction TB
        E["Embed\n(nomic-embed-text)"]
        SIM["Cosine Similarity\n> 0.75"]
        LLM["Local LLM\n(oMLX)"]
        TAG["Auto-Tag"]
        PROP["Proposal\nEvaluator"]
        DB[("ChromaDB")]
    end

    subgraph Output["output/ (Obsidian vault)"]
        SK["skills/\n+ knowledge refs"]
        KN["knowledge/\nshared topics"]
        PR["proposals/\nskill candidates"]
        OB[".obsidian/\nplugin UI"]
    end

    S --> E --> SIM --> LLM --> KN
    S --> SK
    D --> E --> DB --> TAG --> SK
    DB --> PROP --> PR
    DB --> MCP["MCP Server"] --> CC["Claude Code"]
    Output --> OB

Loading

Skills with cosine similarity > 0.75 get their shared knowledge extracted into standalone files. Each skill's frontmatter points back to the knowledge it references.

Documents are chunked, embedded, auto-tagged, and stored in ChromaDB. If a document looks like a reusable coding convention, a skill proposal is generated.

Claude Code connects via MCP to search your knowledge graph while working — passive augmentation with no manual lookups.

Commands

Command	Description
autoskill ingest <dir>	Deduplicate skills, extract shared knowledge
autoskill index <files>	Index documents into knowledge graph
autoskill watch <dir>	Watch for new files, process continuously
autoskill dashboard	Launch web dashboard
autoskill mcp-serve <vault>	Start MCP server for Claude Code
autoskill proposals	List pending skill proposals

Configuration

# .env
LLM_MODEL=gemma-4-26b-a4b-it-4bit        # Any model via oMLX / Ollama / vLLM
LLM_URL=http://127.0.0.1:1111/v1/chat/completions
EMBED_MODEL=nomic-ai/nomic-embed-text-v1.5
SIMILARITY_THRESHOLD=0.75

Architecture

cli.py                 Thin entry point
├── skills/            Skill parsing, dedup pipeline, writer
├── knowledge/         Doc parsing, chunking, ChromaDB, tagging, watcher
├── proposals/         LLM evaluation, progressive threshold tracking
├── mcp/               FastMCP server (7 tools)
├── dashboard/         FastAPI + D3 + Plotly
└── core/              Embedder, LLM client, config, progress

All data stored in .skill-pipeline/ (ChromaDB + JSON). No cloud services required.

Requirements

• Python 3.11+
• An OpenAI-compatible LLM API (oMLX, Ollama, vLLM, etc.)
• ~2GB disk for embedding model on first run

License

MIT