GitHub - kenforthewin/atomic

Source
knowledge-managementknowledge-graphsemantic-searchmarkdownself-hostedmcprustvector-search

Summary

Atomic is a self-hosted personal knowledge base that auto-chunks markdown notes into “atoms,” embeds them for vector search, and links them by semantic similarity. It supports wiki synthesis via LLMs, a spatial canvas for exploration, agentic RAG chat, and exposes an MCP server for integration with Claude and other AI tools. Built in Rust with a React frontend, it runs as a Tauri desktop app, Docker server, or on Fly.io.

Key Insight

  • Architecture is cleanly separated: all business logic in a single Rust crate (atomic-core), with thin transport wrappers for desktop (Tauri), server (actix-web), and MCP (stdio). This makes it easy to self-host or run locally without vendor lock-in.
  • MCP integration is first-class: exposes semantic_search, read_atom, and create_atom tools at /mcp, meaning Claude Desktop or any MCP-compatible client can directly query and write to your knowledge base.
  • Vector search uses sqlite-vec - no external vector DB needed. Everything is SQLite-based, keeping deployment simple.
  • Auto-tagging and wiki synthesis use LLMs (OpenRouter or local Ollama), so you can run fully offline with local models or use cloud providers.
  • RSS feed ingestion and a browser extension for clipping make it a potential replacement for Raindrop + manual insight extraction workflows.
  • iOS app (SwiftUI) and offline-capable browser extension with queue sync add mobile/cross-device coverage.
  • Compared to Obsidian or similar tools, the differentiator is automatic semantic linking and LLM-powered synthesis rather than manual wikilinks.