What it does LlamaFS is a file organizer powered by Llama 3. Point it at a directory and it suggests renames and folder structures based on file contents. It handles text, images (via Moondream), and audio (via Whisper). Two modes: batch for one-shot cleanup, and watch mode for a daemon that intercepts filesystem operations and learns from your manual edits to predict where future files should go.

The interesting bit The watch mode is the clever part: move 1–3 tax documents into a “2023 Taxes” folder and it infers the pattern, then auto-sorts the rest. The authors claim <500ms response times in watch mode by selectively rewriting only changed sections of an index rather than reprocessing everything—plus Groq’s fast inference API.

Key highlights

• Incognito mode routes all requests through local Ollama instead of Groq’s cloud API; same model, no file exfiltration
• Electron frontend lets you preview suggested structures before committing changes
• Image understanding via Moondream, audio via Whisper
• AgentOps integration for latency/cost logging and session replay
• The authors dogfooded it on this repo (“very Meta”)

Caveats

• Groq and AgentOps API keys required out of the box; Ollama switch is labeled “TODO” in the code, not a clean toggle
• Installation is manual clone-and-pip; no packaged release visible
• “Filesystem diffs are hard” — authors’ words, and the roadmap admits this area needs work

Verdict Worth a look if your ~/Downloads is a disaster and you tolerate some setup friction. Skip it if you need battle-tested reliability or can’t risk an LLM misreading a filename.