Obsidian hooks Claude into the note vault: five free repositories and a permissions pitfall

The Obsidian team has open-sourced the set of tools that let Claude work inside your note vault. Obsidian CEO kepano and the official team released five free repositories: obsidian-skills, defuddle, obsidian-importer, obsidian-help, and obsidian-clipper. They solve the same problem: Claude can read and write real Obsidian vault files, instead of stopping at generating a block of Markdown in a chat window. The original tweet is here, and the demo video also comes from the original tweet.

What the five repositories do

obsidian-skills: a collection of Agent Skills maintained by kepano that teaches agents to write Obsidian-style Markdown, Bases, and JSON Canvas, and also bundles Obsidian CLI and defuddle. It is compatible with Claude Code, Codex, and OpenCode. Installation can be done with two commands: /plugin marketplace add kepano/obsidian-skills and /plugin install obsidian@obsidian-skills. In a reply, Ryven suggested installing this first because with two commands, Claude can write real vault files.
defuddle: extracts the main body of a web page into clean Markdown, removing comments, sidebars, and footers to burn fewer tokens. Run the CLI directly with npx defuddle parse <url> --markdown. One commenter said defuddle alone is worth it: the body is clean and there are fewer tokens.
obsidian-importer: migrates historical notes from Notion, Evernote, Apple Notes, Roam, Bear, OneNote, and others into Markdown. If old notes are not migrated in, the agent cannot read them.
obsidian-help: a plain Markdown version of the official documentation, with AGENTS.md, so agents can learn Obsidian from the source.
obsidian-clipper: the official web clipper that saves pages directly into the vault.

The key to a second brain is the loop

Ryven wrote a long tutorial whose title roughly means building a second brain that gets smarter the more you use it with Claude and Obsidian. The structure comes from Karpathy's LLM Wiki idea published in April 2026: don't have the model dig through raw notes every time; first have it organize sources into a wiki, then answer from the wiki. The tutorial divides folders into raw/, wiki/, and output/, plus a CLAUDE.md rulebook. raw is read-only, wiki is written by the model and keeps links, and output holds finished products. Ingest a source, generate summaries and concept pages, and update the index; ask questions using only wiki/ to avoid the model filling in blanks with general knowledge; do a weekly review for contradictions, unlinked pages, and unsourced claims. A common way a second brain dies: only input, no output, until it eventually becomes a dumping ground.

Two costs

The first is tokens. The content of files Claude can read is sent to the model. Pro and Max limits are shared between Claude and Claude Code, and loading hundreds of files for the first time burns through them quickly. The tutorial suggests starting with five to ten files. Don't put confidential material into raw/ unless your plan and work rules allow it.

The second is permissions. Anthropic's documentation says it plainly: CLAUDE.md is context, not enforced configuration. Writing 'never delete files' is only a request. To actually block actions, you need permissions or a PreToolUse hook. Back up your vault, give scheduled tasks the narrowest permissions, and think carefully before enabling an MCP with deletion tools. The tutorial also lists four pitfalls: editing raw/, CLAUDE.md being too long, asking questions that are too open-ended, and not doing maintenance. Each has a corresponding fix.

How to start tonight

Install Obsidian, create a vault, and put three folders in it: raw, wiki, and output. Open this folder with the Code tab in Claude Desktop, let Claude interview you, and generate CLAUDE.md. Pick a source file you really want to keep, drop it into raw/, and let it ingest. Then ask a cross-source question and save the good answer back into wiki/. Ryven's long article has the complete 10 steps and a 30-day plan, with the link below. Don't load hundreds of files at once; first get one source's ingest and query working. Saving ten tutorials is not as good as having Claude write one page of notes with real links.

Original tweet: https://x.com/imryven/status/2118704411333501298

Long-form tutorial: https://x.com/i/article/2115042274614599680