Skip to content

claude-obsidian

Claude Code plugin that files sources into a linked, cited Obsidian wiki and answers from it, based on Karpathy's LLM Wiki idea.

Reviewed by Sheela · Updated

View on GitHub Open source

Our rating

3.8/5

claude-obsidian treats a notes vault like a database, with transactions and citations, which makes it trustworthy and slow to adopt.

What it is

claude-obsidian is a Claude Code plugin and Python CLI that turns sources dropped into an inbox folder into linked, cited Obsidian pages, then answers questions from what the vault already holds. It follows Andrej Karpathy's "LLM Wiki" pattern and files notes in one of four layouts: generic, LYT, PARA or Zettelkasten.

It ships 15 skills and 3 agents, including wiki, wiki-ingest, wiki-query, wiki-lint and a consent-gated autoresearch skill. Every write is shown as a plan and applied only with that plan's hash, inside a journalled transaction that can be rolled back. The same skills run in Codex, OpenCode, Gemini, ZCode, Cursor and Windsurf through a setup script.

Version 3.0.0 briefly renamed the plugin to agricidaniel-obsidian because Claude Code reserves names starting with claude-; v3.0.1 restored the original name the same day.

Our review

This is the most safety-minded way we found to let Claude Code maintain an Obsidian vault, and the safety is also what makes it slow to set up. The repo has about 15,700 stars and shipped v3.0.0 and v3.0.1 on 21 October 2026. The first renamed the plugin because Claude Code now reserves names beginning with claude-; the second restored the old name as a deliberate exception. Its own release notes concede that Claude Code's validator still reports the name as reserved and that claude plugin tag refuses it, so a later Claude Code release could force the rename again. Several projects use similar names; this review covers AgriciDaniel's.

The write path is built like a database's. The vault is picked explicitly, through an environment variable or a .claude-obsidian.json file, and the tool exits without writing if it is unsure. Changes are planned, approved by hash and journalled, and an edit made in Obsidian mid-plan stops the write with a conflict. The README is honest that its lock only coordinates cooperating local writers: synced devices are outside it, so backups still matter.

What it reads is narrower than the pitch. PDFs and EPUBs get metadata and a hash, with no text extraction. Web pages and YouTube become a plan that needs a separate runner, OCR needs an external tool, and native Windows refuses to write, so Windows users need WSL. Since 3.0.0, contextual prefixes no longer fall back to claude -p; without a configured API or local Ollama provider they stay synthetic.

Two open issues point at scale. One asks how a growing corpus fits the hard 16 MB read cap on the source and claim ledgers; another reports Windows session hooks hitting their 5-second timeout. Neither is answered in the docs yet, and independent reviews remain thin, mostly the author's blog and short directory write-ups such as Heatdrop's.

The data story is clean: no telemetry, plain Markdown on disk, and nothing leaves the machine except what Claude Code itself sends to Anthropic. Next to claude-mem, which records coding sessions, this is a research library, and it repays the setup only when the sources you read are the thing you want to keep.

Where it earns its place

  • A PhD student with a few hundred papers already noted in Obsidian

    Ask questions across the literature and get answers cited to sources

    Adopt the existing vault without moving notes, drop new material into inbox/, then run wiki-ingest and wiki-query. PDFs yield metadata only, so text has to be extracted or pasted first; the only cost is the Claude Code plan.

  • A two-person research consultancy sharing a vault over a sync service

    Keep client research linked and free of dead links

    wiki-lint reports dead links, orphan pages and empty sections. The write lock does not cover other devices, so each person needs backups, and an open issue reports phantom folders when transactions run on iCloud Drive under WSL.

  • A knowledge manager at a firm that bans cloud note databases

    Use an AI assistant on internal notes while keeping files on disk

    The vault stays as local Markdown with no telemetry, and web research needs consent per task. Notes Claude reads still go to Anthropic under the firm's Claude plan terms, and Windows staff need WSL to write.

Before you commit

Setup
claude plugin marketplace add AgriciDaniel/claude-obsidian, then claude plugin install claude-obsidian@agricidaniel-claude-obsidian; needs Python 3.11+ and Obsidian for viewing. /claude-obsidian:wiki previews the vault plan and creates it only after approval.
Integrations
Full plugin for Claude Code; the same Agent Skills link into Codex, OpenCode, Gemini, ZCode, Cursor and Windsurf via scripts/setup-multi-agent.sh. Optional Obsidian CLI for search; no MCP server yet (an open feature request).
Data and privacy
Vault is local Markdown; the README states no telemetry and no cloud database. Notes Claude reads are sent to Anthropic as with any Claude Code session; web research and remote models need per-task consent. MIT; no published security advisories.
Cost at scale
No paid edition. Cost is the agent plan, and it scales with how many notes each ingest or query pulls into context rather than with any claude-obsidian fee.
Limits
Metadata only for PDF and EPUB; URLs and OCR need external runners; native Windows is read-only; 16 MB ledger read cap; two plugin renames in one day show the naming issue is unsettled.

What's good

  • Plain Markdown vault you own, no telemetry
  • Planned, hash-approved, reversible writes
  • Claims linked back to stored source copies
  • Web research and remote models off by default

What's not

  • No PDF or EPUB text extraction
  • Native Windows cannot write; needs WSL
  • Plugin name flagged as reserved by Claude Code
  • 16 MB cap on ledgers raised as a scaling issue

Best for

Obsidian users who want Claude Code to build and query a cited research wiki.

Alternatives

claude-obsidian: common questions

Is claude-obsidian worth it?
claude-obsidian treats a notes vault like a database, with transactions and citations, which makes it trustworthy and slow to adopt. We rate it 3.8 out of 5.
How much does claude-obsidian cost?
Free and open source (MIT). You pay for your agent's model usage.
What are the downsides of claude-obsidian?
No PDF or EPUB text extraction; native Windows cannot write; needs WSL; plugin name flagged as reserved by Claude Code; 16 MB cap on ledgers raised as a scaling issue.
Who is claude-obsidian best for?
Obsidian users who want Claude Code to build and query a cited research wiki.
How much does claude-obsidian cost for a team or in production?
No paid edition. Cost is the agent plan, and it scales with how many notes each ingest or query pulls into context rather than with any claude-obsidian fee.
Is claude-obsidian safe for company data?
Vault is local Markdown; the README states no telemetry and no cloud database. Notes Claude reads are sent to Anthropic as with any Claude Code session; web research and remote models need per-task consent. MIT; no published security advisories.
What are the best claude-obsidian alternatives?
The closest alternatives we have reviewed are claude-mem, notebooklm-py and Graphify.

Answers checked Oct 23, 2026.