atomicapp

Self-hosted personal knowledge base that ships the LLM-augmented features as the product, not as something you bolt on top of a markdown vault. Notes ("atoms") get auto-embedded, auto-tagged, and auto-linked on ingest; semantic search, an LLM-synthesized wiki, an agentic chat, and a force-directed spatial canvas all read from the same store. An MCP server is built in so external clients (Claude, Cursor) can search and write atoms directly.

The closest comparison in this vault is the llm-wiki-pattern / obsidian-ai-complete-guide approach, which builds the same outcome from Obsidian + Claude Code skills + a markdown vault on disk. AtomicApp collapses that into one binary with the AI plumbing inside, which is more turnkey but trades away the plain-files-on-disk substrate that makes the Obsidian setup grep-able and editor-agnostic. thinking-space sits between the two: it bundles the same AI features into one app, but keeps markdown-with-frontmatter as the source of truth and treats its index as a rebuildable cache, so the files stay usable outside it.

Architecture

• Backend: Rust + actix-web, SQLite + sqlite-vec for both content and vector embeddings
• Desktop: Tauri v2 wrapping a React 18 + TypeScript + Vite + Tailwind v4 + Zustand frontend
• iOS: SwiftUI app
• Server: same Rust binary, headless, deployable via Docker or Fly.io
• AI providers: BYO via OpenAI-compatible endpoint — OpenRouter, Ollama, or anything that speaks the protocol — used for both embeddings and completion
• MCP: in-process bridge on desktop; HTTP /mcp endpoint with Bearer token auth on remote servers

Atoms are markdown notes with hierarchical tags, source URLs, automatic chunking, and per-chunk vector embeddings. There are no folders to maintain — taxonomy is the tag tree, which the LLM extends as new atoms come in.

What it does

• Semantic search — vector recall, surfaces atoms that share meaning even when keywords don't match
• Wiki synthesis — for a given tag, the LLM writes a synthesized article with inline citations back to the source atoms
• Agentic chat — chat scoped to a tag or the whole library, retrieves mid-conversation, cites atoms instead of hallucinating
• Spatial canvas — force-directed graph view, semantically related atoms cluster
• Auto-tagging — drop in a note, get back a tag tree with topics, people, places, organizations, events
• Browser extension — web clipper that pushes pages straight into the knowledge base
• MCP integration — Claude/Cursor can search, read, and create atoms without leaving their workflow

Watchlist status

On the watchlist because the project is young (repo created 2025-11-25, ~5 months old) and single-author. Counter-signals: 1.3k stars in five months, very active push history (latest commit on the day this page was written), 100+ tagged releases, three deployment shapes already shipped (desktop / server / iOS / browser extension). The thing to watch is whether Ken can sustain the velocity, whether anyone else lands meaningful contributions, and how the data export / migration story holds up — because if the SQLite store ever rots, you can't grep your way out of it the way you can with a plain markdown vault.

Repo

github.com/kenforthewin/atomic — 1.3k stars, MIT, Rust primary, single author.