Tried AgriciDaniel/claude-obsidian: A Self-Organizing Second Brain for Claude Code

AgriciDaniel/claude-obsidian is gaining serious attention on GitHub—up 631 stars today—for a practical reason: it turns Obsidian into an AI-managed knowledge graph without locking your data inside a proprietary platform.

The project follows Karpathy’s LLM Wiki pattern. Drop in a paper, webpage, meeting note, code snippet, or any other source, and Claude Code can read it, identify relationships, create links, and file the result as plain Markdown inside your Obsidian vault. You own the files, the structure, and the ability to migrate them later.

That makes it interesting for developers building a personal knowledge management workflow, an open-source Notion alternative, or a searchable engineering memory that remains inspectable with standard Unix tools.

For teams, I would route the Claude-compatible traffic through an OpenAI/Anthropic-compatible gateway rather than scattering credentials across local machines:

{
  "env": {
    "ANTHROPIC_BASE_URL": "https://b-lost.com/v1",
    "ANTHROPIC_AUTH_TOKEN": "${B_LOST_API_KEY}",
    "ANTHROPIC_MODEL": "claude-fable-5"
  }
}

The same pattern works with Claude Code and other compatible clients, including Cursor, Cline, Roo Code, Windsurf, Aider, NextChat, and LibreChat. A centralized relay gives administrators one place to manage quotas, routing, and audit boundaries while keeping the vault itself local.

For repeated system prompts and large knowledge-graph instructions, native Anthropic /v1/messages prompt caching is especially relevant. B-Lost supports full prompt caching, with cache hits discounted by 90%, which can reduce both latency and cost during iterative note processing. Its relay also applies a 20% discount to official list pricing.

claude-obsidian is not trying to replace your files with a black-box database. Its strongest idea is simpler: let Claude organize the Markdown you already own.