Signal

Build an AI Second Brain with Claude and Obsidian

2 min read

TL;DR

Obsidian holds everything as plain text files you own; Claude sits on top as the brain that reads the vault, files new information, links it, and answers across all of it. This is the LLM Wiki pattern popularized by Andrej Karpathy in April 2026. Because it is all plain text, it is not locked to any one model: point a different model at it next year and it still works.

Actionable items

  1. Keep the split clean: Obsidian is storage, Claude is the brain. Files stay yours locally.
  2. Connect Claude to the vault with the Local REST API community plugin, exposed to Claude as an MCP server (mcp-obsidian, host 127.0.0.1, port 27124). The vault must be open and the plugin enabled for the connection to work. Use only the key after “Bearer”.
  3. Make a root CLAUDE.md profile by having Claude interview you one question at a time (who you are, goals, tone, strengths and weaknesses, projects), then write it to the vault root so context loads every session.
  4. Structure each project as its own folder with a fixed pipeline: Inputs, Process, Outputs, Feedback, plus a project CLAUDE.md naming the one goal and Claude’s role.
  5. Open a single project as its own vault while working so Claude sees only that job. The big vault plans; a single project ships.
  6. Turn any task done more than once into a skill, a saved workflow you trigger by name.
  7. Wire live data through MCP with read-only, scoped access: calendar and Gmail via workspace-mcp.
  8. Put maintenance on autopilot: schedule a daily task that files new inputs, links them, flags stale notes, and writes a short summary of overnight changes.
  9. Community starting points on GitHub: claude-obsidian (AgriciDaniel), obsidian-second-brain (eugeniughelbur, 43 commands, model-agnostic), second-brain-starter (coleam00).

Key quotes

Keys, not prompts. Telling an agent “don’t delete this” is a suggestion, not a safety setting. If it can technically delete a file or send an email, assume one day it will. Control access at the permission level: read-only and scoped keys, not with words.

You own the brain, not the tool.

claude-code-loops · ingest-thread-skill

Graph View

Backlinks