Claude Obsidian is a set of 15 agent skills that turn an Obsidian vault into a knowledge base an agent maintains. It runs in Claude Code as a plugin and in other Agent Skills hosts such as Codex, OpenCode, Gemini, Cursor, and Windsurf. It follows Andrej Karpathy's LLM Wiki pattern: the agent reads each source, writes linked and source-cited pages, and keeps an index current, with no database or embeddings server.
Work follows a loop. Sources enter through a visible inbox and are kept as immutable, content-addressed copies before synthesis. Source and claim ledgers record authority, freshness, support, contradiction, confidence, and review state, and high-risk claims need two independent sources. Ingested material becomes linked pages, indexes, Maps of Content, and Obsidian Canvas views, and later questions are answered read-only from the evidence already in the vault.
The vault stays a normal directory of Markdown, JSON, and source files, and network access (web research, remote models, URL capture) is an explicit opt-in. Every change is one recoverable transaction: parallel workers return drafts only, an orchestrator inspects and applies one bundle under a vault lock with journaled backups, and a changed target is reported as a conflict rather than overwritten.
Retrieval uses deterministic BM25 locally, with optional contextual prefixes and cosine reranking that fall back to BM25 when the model stage cannot be trusted. A mode skill files new notes under Generic, LYT, PARA, or Zettelkasten conventions without moving existing notes. Local files are captured directly, while URLs, YouTube, and OCR need a configured external runner, and PDF and EPUB capture records metadata without semantic extraction.
Features
- Ingest: turns captured sources into linked pages with provenance records
- Query: answers read-only from relevant vault evidence, preferring a grounded refusal over an invented citation
- Lint: reports dead links, orphans, metadata gaps, stale indexes, and empty sections
- Save: stores one scoped answer or insight on request, never an automatic transcript
- Autoresearch: bounded web research with explicit network consent and a separate merge step
- Canvas: creates and maintains Obsidian Canvas knowledge maps
- Retrieval: BM25 with optional contextual prefixes and cosine reranking
- Methodology modes: Generic, LYT, PARA, or Zettelkasten filing for new notes
- Claim ledgers: track support, contradiction, confidence, and review state for each claim
- Transactions: previewed, hash-checked, single-apply writes with journaled backups and recovery
- Obsidian references: skills for Obsidian Flavored Markdown, Bases tables, and a structured review loop
