obsidian-second-brain

obsidian-second-brain is an open-source Claude Code skill built by Eugeniu Ghelbur, an AI Automation Engineer at Single Grain. It connects Anthropic's Claude Code CLI (and three other AI CLIs) to an Obsidian vault, giving Claude the ability to read, write, synthesize, and maintain your personal knowledge base without you opening Obsidian. The project is MIT-licensed, hosted on GitHub, and described by its author as an evolution of Andrey Karpathy's LLM Wiki pattern.

What It Is

obsidian-second-brain is a CLI skill — not an Obsidian plugin — that installs into Claude Code (or Codex CLI, Gemini CLI, or OpenCode) and exposes 43 slash commands for operating an Obsidian vault as a living knowledge OS. Where a traditional vault is a filing cabinet you maintain manually, this skill makes the vault rewrite itself: ingesting a URL, PDF, audio file, or screenshot rewrites 5–15 existing pages rather than appending a new one, contradictions are reconciled automatically, and patterns are synthesized across sources without prompting. The core principle the README calls the "AI-first vault rule" means every note is structured for future-Claude to retrieve and reason over, not for human reading — with machine-readable frontmatter, recency markers per claim, mandatory wikilinks, and confidence levels.

Architecture and Command Layers

The 43 commands are organized into four layers plus always-on background agents:

• Operations (28 commands): save conversations, ingest sources, reconcile contradictions, export snapshots, manage Google Calendar events, log work sessions, and maintain vault health.
• Thinking tools (7 commands): /obsidian-challenge searches your vault for past failures and pushes back on your current idea using your own words; /obsidian-emerge surfaces unnamed patterns from 30 days of notes; /obsidian-connect bridges two unrelated domains; /obsidian-panel convenes distinct perspectives on a decision.
• Context engine (1 command): /obsidian-world loads identity and current state with progressive token budgets so every session picks up where the last one ended.
• Research toolkit (7 commands): pulls live external knowledge via xAI Grok (X posts and trends), Perplexity Sonar (web research with citations), Gemini File Search (vault-grounded synthesis), YouTube transcript API, and podcast RSS/Whisper. /research and /research-deep fall back to free, key-less sources (Wikipedia, HackerNews, arXiv, Reddit, and others) when no Perplexity key is set.
• Always-on agents: a background agent fires after every context compaction; four scheduled agents handle morning briefings, nightly consolidation, weekly reviews, and Sunday vault health audits.

Multi-CLI and Preset Support

The repository ships a build script that compiles the platform-neutral source into four platform-specific outputs. Claude Code uses native slash commands and a CLAUDE.md dispatcher; Codex CLI, Gemini CLI, and OpenCode each get their own dispatcher file and auto-generated routing tables mapping natural-language triggers to command files. The 3 Google Calendar commands are Claude Code–only (they depend on the claude.ai Calendar connector); the other 40 commands are cross-platform. At bootstrap, users choose one of four role presets — executive, builder, creator, or researcher — each generating tailored folder structures, templates, and kanban boards.

Update: v0.10.0 — The Architect (May 2026)

The latest release, v0.10.0 ("The Architect"), published May 31 2026, adds /obsidian-architect: a command that scans a software codebase and writes maintained architecture notes (overview with diagram, per-module notes, key-decisions mined from git history) into the vault. Re-running the command refreshes only content inside generated sentinel markers (<!-- @generated --> blocks) and never overwrites user-added annotations. The release also ships key-less web research (no Perplexity key required), Google Calendar commands, anti-hallucination guards, and a test suite with CI. The project reached 1,534 GitHub stars and 175 forks as of early June 2026, per the repository metadata.

Why It Got Attention

The author's Substack post "I built this for myself. Then 1,374 strangers cloned it." describes the origin: two disconnected tools (Claude and Obsidian) and the institutional-amnesia problem of AI sessions that start from scratch every time. The README frames the project as a direct extension of Karpathy's LLM Wiki gist, adding self-rewriting pages, automatic contradiction reconciliation, scheduled agents, and an AI-first note format that the original pattern lacks. A first community fork, scholarbrain, targets academic research use cases, and the ECOSYSTEM.md file documents a pattern for domain-specific forks.