Your agent can read Markdown. That does not mean it can edit an Obsidian vault without breaking it.
Wikilinks become dead links. Callouts lose their syntax. Bases files turn into invalid YAML soup. JSON Canvas graphs lose node IDs. The vault still opens, but the graph view lies to you.
kepano/obsidian-skills is Steph Ango's answer: official Agent Skills for Obsidian formats, MIT licensed, following the agentskills.io standard so Claude Code, Codex, and OpenCode can share the same packs.
This is not a chatbot sidebar inside Obsidian. It is teaching your coding agent the file formats your knowledge base already uses.
The five skills (and what each fixes)
| Skill | Format | Agent capability |
|---|---|---|
obsidian-markdown | .md | Wikilinks [[note]], embeds ![[file]], callouts, properties, tags |
obsidian-bases | .base | Database views, filters, formulas, summaries |
json-canvas | .canvas | Spatial graphs with nodes, edges, groups |
obsidian-cli | CLI / URI | Vault automation, plugin and theme dev workflows |
defuddle | Web → MD | Clean markdown from noisy web pages via Defuddle |
Each skill is a folder with SKILL.md plus optional references. Progressive disclosure keeps token burn low: the agent loads full instructions only when the task needs that format.
If you already use the open Agent Skills format in client repos, obsidian-skills is the vendor-maintained pack for Obsidian specifically.
Install paths (pick your agent)
Claude Code (plugin marketplace):
/plugin marketplace add kepano/obsidian-skills /plugin install obsidian@obsidian-skills
Codex CLI:
Copy the skills/ directory into ~/.codex/skills/ per the Agent Skills spec.
OpenCode:
Drop skills under ~/.opencode/skills/. OpenCode auto-discovers SKILL.md files without extra config.
Vault-local Claude workflow:
Some teams copy skill folders into .claude/ at the vault root so project context and Obsidian skills travel together.
Why the Obsidian CEO shipping skills matters
Tool vendors usually ignore agent integration until a startup builds a flaky MCP wrapper.
Kepano publishing MIT skills signals that open Agent Skills are the integration layer for knowledge work, not proprietary plugin silos. Obsidian's philosophy is files over apps. Agent Skills are instructions over opaque memory databases.
That aligns with how I already run client knowledge bases:
- Plain text canonical store (Obsidian or git-backed Markdown)
- Agents that read/write files with provenance
- No second memory system that drifts from repo truth
obsidian-skills does not replace your retrieval architecture. It stops agents from corrupting the canonical store while they work.
Practical workflows I would wire this week
Meeting notes → vault. Paste a transcript, invoke obsidian-markdown skill context, get valid callouts and wikilinks to people and project notes.
Research capture. Run defuddle on a noisy article, save clean Markdown with properties, link from a daily note.
Project dashboards. Ask the agent to extend a .base view with filters your PM actually uses, not a CSV export hack.
Canvas planning. Generate JSON Canvas maps for architecture discussions that stay editable in Obsidian's graph UI.
CLI batch ops. Rename, move, and tag notes through obsidian-cli skill instead of brittle shell scripts on raw files.
Limits and honest caveats
- Skills teach format compliance, not your folder taxonomy. You still need VAULT_RULES or equivalent onboarding docs.
- obsidian-cli skill assumes Obsidian CLI availability. Verify against your Obsidian version and platform.
- Agent Skills do not solve embedding search. Pair with grep, ripgrep, or your RAG layer as needed.
- Highly custom plugins (Dataview JS, exotic templates) may still need bespoke skills you author.
Test on a vault branch or git worktree before letting an agent bulk-edit production notes.
Relationship to other vault-agent projects
The digest ecosystem also buzzed about agentcairn, Franklin, and memory vault MCPs. Different layer:
| Layer | Example | Job |
|---|---|---|
| Format skills | obsidian-skills | Valid Obsidian file syntax |
| Memory systems | agentcairn, agentkeep | Cross-session recall and provenance |
| In-app agents | Pivi plugin | Chat UI inside Obsidian |
Start with obsidian-skills if your pain is broken wikilinks and mangled callouts. Add memory tooling when cross-session recall becomes the bottleneck.
If you want an Obsidian-backed agent workflow for ops docs, client delivery, or personal PKM without vault corruption, book a free discovery call.