A summary of @undefinedKi’s guide on wiring Claude to an Obsidian vault so it acts as a persistent, self-organizing “second brain.” It’s an applied version of Andrej Karpathy’s LLM Wiki pattern (April 2026): keep everything as plain text, and let a model read, link, and grow it for you.
The architecture
Two tools, two jobs:
- Obsidian is the storage. Plain-text notes on your own machine that link to each other and form a graph. Because it’s just files, the brain is yours and portable.
- Claude is the brain on top. It reads the whole vault, files new information where it belongs, links it to what’s already there, and answers questions across all of it.
The key property: since the store is plain text, it’s model-agnostic — point a different model at the same vault next year and it still works. You own the brain, not the tool.
How it connects
Claude reaches the vault through MCP (the standard protocol Claude uses to talk to other apps). Obsidian exposes a door via the Local REST API community plugin; you register that endpoint with Claude Code using a scoped API key. The vault is only reachable while Obsidian is running.
The layers that make it “smart”
The guide’s real content isn’t the wiring — it’s the structure that keeps the system focused and self-maintaining:
CLAUDE.mdat the vault root is the persistent context layer — who you are, your goals, how you want to be talked to. Loaded every session, so you never re-explain yourself. (Built by having Claude interview you rather than typing it out.)- Projects are folders, each with
Inputs / Process / Outputs / Feedbackand their ownCLAUDE.md. Ideas land in Inputs, work happens in Process, results in Feedback. - Open a single project as its own vault so Claude sees only that job. “The big vault plans; a single project ships.”
- Skills are saved markdown workflows triggered on command — anything you do more than once becomes one.
- Live data via MCP (calendar, email, Slack, Notion), granted read-only.
- Scheduled tasks put it on autopilot: a daily job that files new notes, links them, flags stale ones, and summarizes what changed overnight.
The one principle worth keeping
Keys, not prompts. Telling an agent “don’t delete this” is a suggestion, not a safety control. If it can delete a file or send an email, assume one day it will. Enforce limits at the permission level — read-only, scoped keys — never with wording in a prompt. This is the most transferable idea in the piece and applies to any agent setup, not just this one.
Pointers
Ready-made repos that scaffold the whole thing (all plain text + scripts): claude-obsidian (AgriciDaniel), obsidian-second-brain (eugeniughelbur), and second-brain-starter (coleam00).
Takeaway
The reframe: you’re not building a Claude setup, you’re building your own memory that compounds. Same subscription, different machine — persistent context and a graph that maintains itself, instead of a cold chat box every session.
Source: @undefinedKi on X, 2026-06-20