Most note-AI workflows lose the raw evidence or scramble provenance; that makes retrieval brittle and claims hard to audit. claude-obsidian treats the vault as first-class data: it captures immutable source copies, compiles them into linked Markdown pages, and surfaces provenance so every important claim can be traced back to original evidence. The result is a compounding, queryable knowledge base you own rather than a transient model context.
What Sets It Apart
- Content-first, not model-first — it preserves content-addressed copies of sources before any synthesis, so summaries always point back to verifiable evidence (so what? you can audit and correct claims instead of trusting a transient LLM context).
- Provenance and transaction safety — operations are planned, inspected, and applied as recoverable transactions with SHA-256 plans and journaling (so what? parallel workers produce drafts, but only one inspected bundle changes the vault, preventing silent overwrites and race conditions).
- Vault-native outputs — generated notes are plain Obsidian Markdown, Canvas and Bases-friendly, and live in a normal filesystem vault (so what? your knowledge remains portable, searchable, and usable without a proprietary database).
- Retrieval-first tooling — BM25 local retrieval, optional reranking, and explicit egress gates for model-based steps enable grounded Q&A workflows (so what? answers come from the vault evidence, reducing hallucination risk and making reviewable research reproducible).
Who it's for and tradeoffs
Great fit if you want an auditable, compounding personal knowledge base where an LLM does the filing rather than replacing your files. It's aimed at researchers, knowledge workers, and teams that need source-grounded summaries, evidence-led rollups, and explicit upgrade/migration controls.
Look elsewhere if you need a plug-and-play cloud sync note app or a simple transcript recorder — claude-obsidian deliberately avoids hidden caches and automatic uploads. It also assumes an Obsidian workflow, Python 3.11+ tooling, and occasional external runners for URL/YouTube/OCR tasks, so users seeking purely web UI or zero-setup hosted services may find it heavyweight.
Where it fits
This isn't a competitor to general-purpose hosted note apps; it's a local-first bridge between LLM-enabled workflows (Claude Code / Agent Skills) and durable, provenance-aware PKM. Think of it as the layer that makes RAG-style research auditable and maintainable inside an Obsidian vault.
