// Skill profile
AgriciDaniel/claude-obsidian is a local-first knowledge management system that turns source materials into interconnected Obsidian notes using Claude Code. Based on Karpathy's LLM Wiki pattern, it operates through a capture-ground-connect-use cycle: sources enter through an inbox as immutable, content-addressed copies; claims are linked to source evidence via ledgers; notes are automatically linked and indexed; and the vault answers queries from existing evidence rather than generating from scratch. The system ships 15 coordinated skills covering ingestion, querying, linting, research, and canvas creation, with four methodology modes (Generic, LYT, PARA, Zettelkasten) for organizing notes. Transaction-based safety provides atomic writes, rollback capability, and operation recovery. All output is plain Markdown — the vault remains portable and readable without the system. Install by cloning the repo and running the init script, or use the non-destructive adopt workflow for existing vaults. 13.6K GitHub stars, 1,412 forks, MIT-licensed.
Claude Obsidian is a self-organizing AI second brain that bridges Claude Code and Obsidian into a connected knowledge management system. Based on Karpathy’s LLM Wiki pattern, it takes the approach that AI should organize and connect your knowledge rather than just answer questions about it. Drop any source material — papers, articles, documentation, notes — and Claude reads, links, and files it into a growing knowledge graph of plain Markdown files you fully own. The vault becomes more useful over time as connections compound deliberately through evidence-grounded linking rather than through automatic transcript dumping.
Capture-ground-connect-use cycle: The system follows a disciplined information pipeline. Sources enter through an inbox and are stored as immutable, content-addressed copies — the original material is always preserved and retrievable. Claims extracted from sources are linked to their evidence via ledgers, maintaining provenance. Notes are automatically linked and indexed, building a knowledge graph. When you query the vault, answers come from existing evidence rather than being generated fresh, which means responses are grounded in what you have actually read and stored.
15 coordinated skills: Claude Obsidian ships a comprehensive skill set covering the full knowledge management workflow — ingestion of new sources, querying existing knowledge, vault linting and health checks, research workflows, and canvas creation for visual exploration. The skills coordinate to maintain vault consistency rather than operating as independent tools.
Four methodology modes: The system supports Generic, LYT (Linking Your Thinking), PARA (Projects, Areas, Resources, Archive), and Zettelkasten organizational methodologies. Each mode structures the vault according to its respective principles, so users can work within their preferred PKM framework rather than adapting to a new one.
Transaction-based safety: All vault modifications use atomic writes with rollback capability and operation recovery. If a complex multi-file operation fails partway through, the vault returns to its previous consistent state. Operations that the system cannot safely perform fail clearly with an explanation rather than being silently approximated.
Researchers use Claude Obsidian to build a literature review vault — feed it papers and it extracts claims, links them to sources, and builds Maps of Content that reveal connections across the literature. Developers use it as a project knowledge base — architecture decisions, meeting notes, and technical references are automatically cross-linked. Lifelong learners use it as a personal wiki where sources are preserved alongside synthesized understanding. Teams migrating from Notion or other tools use the adopt workflow to bring an existing vault under Claude Obsidian’s organizational system without restructuring what they already have.
Claude Obsidian is local-first — all data stays on your machine in plain Markdown files. This means full ownership and portability, but also means the knowledge graph is per-device unless you sync the vault through Obsidian Sync, iCloud, or git. The system requires Claude Code as the primary agent (other agents are supported through portable skill links but with reduced capability). The init process creates a structured vault layout, so users with heavily customized existing vaults should use the adopt workflow rather than init.
Knowledge workers who want their AI to organize information rather than just retrieve it. Obsidian users who want automated linking and indexing without losing control over their vault structure. Researchers who need evidence-grounded answers from their own collected materials. Anyone following the Karpathy LLM Wiki pattern who wants a production-ready implementation rather than a DIY setup.
AI agents that work well with Claude Obsidian.
Universal memory layer for AI agents — remembers user preferences, past interactions, and context across sessions.
Vectorless, reasoning-based RAG — PageIndex builds a hierarchical tree index over long documents and uses LLM reasoning to retrieve, instead of cosine similarity over embeddings.