Every download filed the moment it lands, and every move one click from undone.
A local Gemma agent for your Downloads folder, with a small robot named Zee.
Quick start · Zee · Safety · Your rules · Benchmarks · How it works
Receipts, booking confirmations, photos, manuals, and half-finished downloads pile up, because opening, renaming, and moving each file by hand is tedious. Dropzone Zero reads each new download with Gemma 4 running on your Mac, decides what it is, names it, and files it. Risky moves wait for you, in Zee or on your phone. Every move is hash-checked and journaled, so one click puts any of it back, byte for byte.
scan0001.pdf -> Finance/Receipts/2026-10/2026-10-08-metro-cab-co-trip-receipt.pdf
IMG_2041.jpg -> Finance/Receipts/2026-10/2026-10-11-juniper-market-receipt.jpg (a photo of a receipt)
invite.ics -> Calendar/2026-10-29-local-open-source-meetup-invite.ics
Free and MIT-licensed. The core needs no account or API key, and your files never leave the machine.
Built for the Hacktoberfest 2026 DEV Challenge (Week 3, "Slow and Steady").
| 🏠 Your files stay on your Mac | Gemma 4 (gemma4:e4b) reads text, PDFs, Office files, and photos through Ollama, locally. Cloud services only ever see synthetic test data, a label like "move 1 installer", or counts. |
| 🧾 The model reads; code files | Gemma answers in strict JSON: a kind, the date printed in the document, a title. Code builds the name and folder, and a date stays only if it is really written in the file. |
| 🛑 Allow, ask, or block | A guard judges every move before it runs. Installers and unsure guesses wait for you; moves out of the folder, into hidden folders, or that change a file's type never run. |
| ↩️ Undo is exact | Each file moves only if its SHA-256 still matches the preview, and is hashed again after. Every rename is journaled, so dz undo or Undo last restores it. Nothing is ever deleted. |
| 🤖 No terminal needed | Zee sits in a corner of the screen, shows what it is doing, asks with Move it / Leave it, and is the only thing you launch. |
| 📱 Asks on your phone | With Auth0 CIBA, the same question arrives as a Guardian push. Only the kind of move leaves the machine. |
| 🗂 Your own rules | A .dropzonerules.json sets where each kind of file goes and how it is named, and the guard still checks every move. |
| 🔎 Find it by describing it | "photo of a receipt from a market" finds the right file, using Gemma's image descriptions and EmbeddingGemma, all local. |
 |
 |
| The CLI: preview, run, and undo on synthetic downloads | Zee hiding: it dives into a hole, and pops back out when called |
 |
 |
| Phone approval: the real push for a synthetic installer (account covered) | Tracing: one span per Gemma call, tokens and timing, never the prompt |
Requires Node.js 22.9 or newer, pnpm, and Ollama for the local model. The optional review gate also needs uv.
pnpm install
ollama pull gemma4:e4b # about 6.6 GB, once
ollama serve # skip if the Ollama app is already running
pnpm seed # creates demo/inbox with 31 synthetic files
pnpm seed:history # optional: synthetic past decisions for the review gate
pnpm dz preview demo/inbox
pnpm dz run demo/inbox # asks before moving anything
pnpm dz undo demo/inbox # puts every file backOn an Apple Silicon laptop the 31-file demo preview takes about 70 to 100 seconds with gemma4:e4b, most of it spent
looking at images. On a real Downloads folder with larger PDFs and photos, expect about 5 seconds per file. Gemma 4
thinks before answering by default; Dropzone Zero turns that off, because for one-file classification it was about 15
times slower with no better answers.
For the desktop app, which starts Ollama and the background daemon for you:
pnpm app:build # builds Zee.app (about 10 MB) with Tauri
pnpm dz autostart enable ~/Downloads # Zee starts now and at every loginAll commands: Commands and everyday use.
| What | Result | How |
|---|---|---|
| Classifier, 63 synthetic test files | 98.4% kind, 100% date, 0% invalid, p50 3.6 s | dz eval, local gemma4:e4b |
| Qwen3.5-4B fine-tuned on Tinker (a control) | 100% kind, date 92.1% → 100% after LoRA | ml/finetune.py, scored by dz eval --provider replay |
| Korean titles (the first eval run's bug) | 85.7% → 98.4% after fixing ASCII-only names | dz eval |
| Search, 7 test questions on the demo folder | 7/7 put the right file first | dz find |
| Receipts to CSV, demo folder | 6/6 receipts and invoices right | dz receipts |
| A run killed after 15 of 30 files | 30 moved, 0 failed after dz worker resumed it |
dz run --temporal |
Every test file is synthetic, and the set was written alongside the app, so treat these as regression checks, not accuracy claims. Three models landing between 98% and 100% also means the set is too easy to tell good models apart. Details: Measuring the classifier.
new download
→ rules: still downloading? → skip · empty, duplicate (same SHA-256), temp? → quarantine
→ Gemma 4 (Ollama, local) reads it → strict JSON: kind, printed date, title, confidence
→ code builds the name and folder (defaults or .dropzonerules.json); a date stays only if printed in the file
→ guard: ALLOW · REVIEW (Zee, terminal, or phone) · BLOCK + TabPFN: "will you undo this?"
→ harness: judge again, check hash, rename, hash again, journal → undo from the journal
- The model proposes, the guard decides. No file name or document text is ever treated as an instruction; the worst a hostile document can do is get itself misfiled.
- The harness trusts nothing before it. It re-checks every move against the disk, never overwrites, follows no symbolic links, and refuses a plan edited by anything but Dropzone Zero.
- It learns what you undo. Single-file undos teach a local TabPFN review gate, from metadata only, which moves to ask about next time.
- Private by default. Telemetry is off unless you set
SENTRY_DSN, and even then file names, paths, prompts, and contents are never sent.
Deep dive: How it works and Safety.
| Guide | What's in it |
|---|---|
| Commands | Every dz command, dz tidy (Mastra), watching a folder, start at login, durable runs with Temporal |
| Safety | ALLOW / REVIEW / BLOCK, how a run stays safe, what each part can touch, .dropzoneignore, the TabPFN review gate |
| Your rules | .dropzonerules.json: where each kind of file goes and how it is named |
| Zee | The desktop companion, what its faces mean, approving from your phone with Auth0 |
| Search | Finding a file by describing it, with Gemma captions and EmbeddingGemma |
| Follow-ups | Calendar events for invoices, bookings, and warranties; receipts to CSV |
| How it works | How a plan is made, settings, the classifier eval and the Tinker control, telemetry, demo data, roadmap |
Does anything leave my machine?
Not your files. Reading, describing, and searching all run on Ollama locally. The optional parts send only what the table in Safety lists: a label such as "move 1 installer" for a phone approval, a sentence of kinds and counts for the voice, counts and timings for telemetry.
Will it delete anything?
No. Duplicates and temp files go to .dropzone/quarantine inside the folder, which only you empty. Everything else
is a rename, and undo puts it back.
What if it files something in the wrong place?
Click Undo last in Zee, or run dz undo. dz undo --file <name> puts back a single file and teaches the review
gate that this kind of move was wrong.
Can I choose the folders and names?
Yes. List only the kinds you want to change in a .dropzonerules.json; see Your rules. A rule
that points outside the folder or into a hidden one stops the run with a message.
Does it need a GPU or the internet?
No GPU beyond what Ollama uses on your Mac, and no internet after the one-time model download.
- macOS only, so far. Zee, start at login, notifications, and calendar removal use macOS features, and that is the only system it has been tested on.
- Scans with no text layer are classified by name and size only, and a photographed invoice gets no due-date reminder, because due dates are read from text.
- The test set is too easy to separate good models; it needs messier documents before fine-tuning Gemma itself makes sense.
MIT. Zee bundles three.js (MIT).
Built with PriorLabs-TabPFN. The review gate uses the TabPFN v2 weights under the Prior Labs License.