Best AI Agent Skills for Obsidian in 2026

Obsidian has the best agent story of any note app right now, and it's for an unsexy reason: the vault is a folder of Markdown files on your disk. No API to authenticate against, no export step, no proprietary blob. An agent that can read and write files can already use your notes.

That makes the skills in this category unusually good. It also means you should be clear about what you're actually buying, because two very different things get labeled "Obsidian AI."

The six skills

Two verified, four community, all free. The star counts span two orders of magnitude, and the gap tells you something real about the top entry.

obsidian-skills: start here

Install: claude skill add kepano/obsidian-skills

kepano is Obsidian's CEO, and this is the closest thing to a first-party agent integration the app has. It teaches your agent to use Markdown, Bases, JSON Canvas, and the Obsidian CLI. Verified tier, last updated 2026-06-08, compatible with Claude Code and Codex CLI.

The 43,957 stars deserve a caveat: that number reflects the repository's overall standing, not this skill's install base in isolation. Weight the authorship more than the count.

What makes it the right first install is the approach. It doesn't run a server between your agent and your vault. It teaches the agent the formats your vault already uses, including Bases and JSON Canvas, which most third-party integrations ignore entirely. Fewer moving parts, nothing to keep running, and it degrades gracefully because your notes stay plain files.

Limitation: no semantic search. The agent reads what you point it at and what it can find by name or grep. For a 300-note vault that's fine. For 5,000 notes where you can't remember what you called things, you'll want one of the MCP options alongside it.

mcp-obsidian: the established MCP route

Install: claude mcp add mcp-obsidian -- npx -y MarkusPfundstein/mcp-obsidian

4,249 stars, verified tier, updated 2026-05-15, working through Obsidian's REST API community plugin. Claude Code, Cursor, and VS Code Copilot.

The REST API dependency is the thing to understand. This talks to a running Obsidian instance rather than to files on disk, which means the app has to be open for the agent to work. In exchange you get the live application's view of the vault: resolved links, plugin state, the actual graph rather than raw text.

Whether that trade is worth it depends on how you work. If Obsidian is always open on your desktop, it costs you nothing. If you want an agent touching your vault on a server at 3am, this is the wrong architecture and obsidian-skills or a file-based approach fits better.

obsidian-second-brain: the ambitious one

Install: claude skill add eugeniughelbur/obsidian-second-brain

3,802 stars, community tier, updated 2026-08-01, making it the most recently maintained skill on this list. It turns a vault into what the author calls a living second brain: autonomous writes, thinking tools, knowledge ingestion, scheduled agents, and a _CLAUDE.md for cross-surface context.

This is the one I'd be most careful with, and also the one doing the most interesting work. "Autonomous writes" and "scheduled agents" mean an agent modifying your notes without you watching. That's the actual promise of an AI second brain, and it's also how you end up with 40 files you didn't write and can't distinguish from your own thinking.

If you install it, do two things first. Put the vault under version control: git init in the vault folder, commit before you start, and you can diff or roll back anything. And point it at a subfolder rather than the whole vault until you trust its output. Both take five minutes and turn an irreversible experiment into a reversible one.

obsidian-mcp-tools: semantic search

Install: claude mcp add obsidian-mcp-tools -- npx -y jacksteamdev/obsidian-mcp-tools

833 stars, community tier, updated 2026-05-13. Adds semantic search and custom Templater prompts to Claude or any MCP client, across Claude Code, Cursor, and VS Code Copilot.

Semantic search is the feature that changes what an agent can do with a large vault. Keyword search finds notes containing the word you typed; semantic search finds the note about the same idea under a different name, which is the actual retrieval problem in personal knowledge management. If your vault has grown past the point where you remember your own filenames, this is the gap-filler.

The Templater integration is narrower but useful if you already run Templater, because your existing note templates become things the agent can invoke rather than approximate.

obsidian-mcp-server: the thorough one

Install: claude mcp add obsidian-mcp-server -- npx -y cyanheads/obsidian-mcp-server

646 stars, community tier, updated 2026-08-02, the freshest timestamp here. It provides reading, writing, searching, and management of notes, tags, and frontmatter.

Frontmatter handling is the distinguishing feature. If you use YAML properties for structure, such as status fields, dates, tags, and relations, a skill that treats frontmatter as data rather than as text at the top of a file is meaningfully more capable. Anyone running Dataview or Bases queries over their properties should look here first.

obsidian-plugin-skill: for developers only

Install: claude skill add gapmiss/obsidian-plugin-skill

180 stars, community tier, updated 2026-07-30. This one is for building Obsidian plugins, not for using a vault. Different audience entirely, included so nobody installs it expecting note management.

Two architectures, pick deliberately

The split running through this list matters more than any individual feature.

File-based (obsidian-skills, obsidian-second-brain) treats the vault as what it is, Markdown on disk. Works with Obsidian closed. Works over SSH. Works in a scheduled job. Version-controls cleanly. Can't see anything Obsidian computes at runtime.

Server-based (the three MCP entries) talks to a running Obsidian. Sees resolved links, plugin state, semantic indexes. Requires the app open, and adds a process that can break independently of both your agent and your notes.

Most people should start file-based with obsidian-skills, then add an MCP server when they hit a specific wall, usually semantic search on a big vault or frontmatter-aware queries.

A rough threshold from running both: under about 500 notes, file-based alone is fine and the MCP server is overhead you'll notice more than benefit from. Past a few thousand notes, retrieval becomes the binding constraint and semantic search stops being optional. The middle range depends less on note count than on whether your naming is consistent enough for an agent to find things by filename.

A word on trust

Four of six are community tier, and all six get read-write access to your personal notes. That's a different sensitivity profile than a code repository. Your vault probably contains drafts you wouldn't publish, notes about people, and things you wrote assuming nobody would read them.

Before installing any community-tier skill here: read what it does with vault contents, check whether anything is sent off-machine, and commit your vault to git first. The specific thing to grep for is any HTTP client pointed somewhere other than your own machine, since a vault skill has no legitimate reason to make outbound calls unless it advertises a sync or embedding feature. If it does advertise one, find out where the embeddings are computed before you index 5,000 personal notes into somebody else's service. Our writeup on reducing the risk of community skills goes deeper on the audit process.

If you're building the vault side rather than the agent side, how to build a second brain with AI tools and our knowledge management roundup cover that. There's also a step-by-step Obsidian MCP install guide. For the wider stack, see our lists for researchers and writers.

FAQ

Which Obsidian skill should I install first? obsidian-skills from kepano. It's authored by Obsidian's CEO, verified tier, and needs no server running. Add an MCP option later if you need semantic search or frontmatter queries.

Do I need the Obsidian app open for these to work? For mcp-obsidian, yes, because it goes through the Obsidian REST API plugin, which requires a running instance. obsidian-skills and obsidian-second-brain work against files on disk, so the app can be closed.

Can an agent overwrite my notes? Yes, and obsidian-second-brain does so by design. Run git init in your vault and commit before installing anything with write access. Every change becomes reviewable and reversible.

What's the difference between an Obsidian skill and an Obsidian MCP server? A skill teaches the agent knowledge, such as formats, conventions, and CLI commands, with nothing running in the background. An MCP server is a live process the agent calls to reach your vault. Skills are simpler and more portable; servers can do things that need a running application.

Is any of this free? All six are free and open source. You need Obsidian itself, which is free for personal use, and mcp-obsidian additionally requires the Local REST API community plugin.