Tidy

Tell your Mac what to do with your files — in plain English.
Review exactly what will happen. Undo anything.

Download · Try it in 2 minutes · How it works · How AI is used · Benchmarks · Ideas

---

Why Tidy?

Your Mac's storage is a mystery, organizing thousands of files by hand takes an afternoon, and nobody trusts a bot with rm -rf.

Tidy fixes all three. Type what you want. Tidy shows exactly which files and folders it would touch, with sizes. You approve. Everything goes to the Trash with a recorded location, so any change can be undone from inside the app.

🗣️ Plain English “Delete the 10 biggest files”, “organize by date”, “remove the instances that aren't the 26.2 version”. No rules to learn.

🛡️ Safe by construction Opt-in folders, review-then-approve, no overwrites, no permanent deletes, a durable journal and one-click Put back.

🔒 Fully local No account, no cloud, no telemetry. The optional AI model runs on your Mac; your files never leave it.

✨ Features

💬 Chat that actually does things Create, move, rename, delete, sort, de-duplicate, change extensions — with folder-name fuzziness (“life and hell” → liveandhell-template), lists (“delete A, B and C”), multi-step requests (“rename X to Y and move it into Z”, “create a folder Invoices and move the PDFs into it”) and follow-ups (“delete all of them”).	💽 A disk explorer that tells the truth Every folder on your Mac with its real allocated size, cached so it opens instantly and refreshed in the background. Drill in, sort by size/name/date, open any file's location, trash it — and see why 800 GB are used.
🕘 History with Put back Every change is journaled: where it was, where it went, where it is in the Trash. Undo a batch of moves, or put a trashed folder back — Tidy refuses to overwrite anything.	🧠 Bring your own model Instant deterministic engine for everyday requests; an optional local model (Qwen3 by default) for unusual ones. Paste any Hugging Face .gguf link to add your own, pinned by checksum.

More: a tiny always-on-top assistant, light & dark, and everything reachable from the keyboard

  

The shrink button turns Tidy into a 360×580 assistant pinned to a screen corner (Esc to expand). All screens stay reachable from its icon bar.

🎬 See it in action

You say	Tidy does
“What's taking the most space?”	Cards for every top-level folder with sizes and bars; click one to open it.
“Remove the curseforge instances that are not the 26.2 version”	Reads each instance's own version file, keeps the matches, lists the rest whole.
“Organize this folder by type”	Images, Documents, Installers… in reviewed 500-item batches; Undo all restores it.
“Rename setup.dmg to Installer and move it into Archive”	Two steps, one reviewed plan and one approval; Put back undoes both at once.	—
“Replace spaces with underscores in file names”	Bulk renames with collision checks; contents never change.
“Find duplicate files and remove the extra copies”	Hashes files of equal size, keeps the oldest, trashes the rest.
“List all my projects”	Finds them by package.json, Cargo.toml, .git… with type and last change.

Every row above is covered by an automated end-to-end test on a mock folder — see Quality.

🛡️ Safety is the product

flowchart LR
    A["💬 You ask"] --> B["🧭 Request engine<br/>(deterministic, optional local AI)"]
    B --> C["📋 Reviewable plan<br/>exact files, sizes"]
    C -->|you approve| D["🔐 Safety engine<br/>validate · fingerprint · journal"]
    D --> E["📦 Execute<br/>no-replace moves · native Trash"]
    E --> F["✅ Verify"]
    F --> G["🕘 History<br/>Undo · Put back"]
    B -. "never touches files" .- X(("AI"))

Loading

Guarantee	How
You choose what Tidy can see	Per-folder switches; everything else is invisible to the assistant.
Nothing changes without approval	One-use, expiring tokens bound to the exact actions; a folder edited after review refuses to run.
No overwrites	RENAME_EXCL no-replace moves; collisions are skipped and reported.
No permanent deletion — anywhere in the code	Removal is the native Trash, with its recovery location recorded for Put back.
Symlink & race safe	Handle-relative, no-follow traversal; mount boundaries respected; .git internals never opened.
Only one crate can mutate files	crates/safety. The UI, the request engine and the AI model can only propose.
Offline	The only network use is you pressing Download model (pinned by revision + SHA-256).

Full threat model: docs/ARCHITECTURE.md.

🤖 AI, used where it helps

flowchart TD
    Q["Your request"] --> E{"Deterministic engine<br/>understands it?"}
    E -->|"yes · ~1–75 ms"| P["Reviewed plan"]
    E -->|no| M["Local model (optional)"]
    M --> R1["① Restate as a plain command<br/>using your real folder names"] --> E
    M --> R2["② Build folders from<br/>grammar-constrained matching rules"] --> P
    M --> R3["③ Answer from your index"]

Loading

Small models are unreliable planners but good at rewriting and classifying, so Tidy uses AI for that and code for precision. Details, model list and how to add your own: docs/HOW-AI-IS-USED.md.

📦 Install

Download Tidy-<version>-macos-arm64.dmg from Releases, drag Tidy to Applications. Builds are not notarized yet, so on first launch right-click → Open, or:

xattr -dr com.apple.quarantine /Applications/Tidy.app

Check the download: shasum -a 256 -c SHA256SUMS.txt. Requires macOS 11+ (Apple Silicon recommended for local AI; Metal is used automatically).

Build from source

Needs Rust 1.90+, Node 22+, Xcode command-line tools and CMake.

git clone https://github.com/NicolasLasch/Tidy.git && cd Tidy/apps/desktop
npm ci
npm run worker:build     # pinned llama.cpp worker (once, ~5 min)
npm run desktop          # development app with hot reload
npm run package          # release bundle → target/release/bundle/macos/Tidy.app

🧪 Try it in 2 minutes

Never point a new tool at your real Documents first. Generate ~1,000 deliberately messy files:

cargo run --example make_mock -p tidy-agent-runtime -- ~/TidyMock --large

Open Tidy → Storage → Add folder → pick ~/TidyMock/Inbox, then ask:

1. what's taking the most space?
2. organize this folder by type → approve → organize by date → History → Undo all
3. delete duplicate files
4. replace spaces with underscores in file names
5. change all .txt files to .md

🏁 Quality

• 140+ automated tests: safety refusals, indexer, request engine, and end-to-end scenarios that run the real safety engine on real files — including a 1,000-file messy tree (organize by type/date then undo everything, bulk renames, extension changes, de-duplication, 500-action batches, Put back) verified byte-for-byte. CI runs fmt, clippy -D warnings, the full suite and a secret scan on every push.
• Fast: understanding a request takes ~1 ms at 1,000 files and ~30–75 ms at 50,000; a 50,000-file folder moves to the Trash (and back) with one atomic rename.

1,000 files	10,000 files	50,000 files
request 1 ms · rescan 45 ms	request 7 ms · rescan 0.46 s	request 31 ms · rescan 3.0 s

Full tables: docs/BENCHMARKS.md · reproduce with cargo run --release --example benchmark -p tidy-agent-runtime.

🧱 Under the hood

Tauri 2 (Rust) + React/TypeScript · SQLite/FTS5 index · llama.cpp worker (Metal/CPU) · Qwen3 models via Hugging Face · no Electron, no Python, no cloud.

crates/platform       authorized roots, no-follow handles, protected paths
crates/file-indexer   scanner + SQLite index
crates/organization   pure planning (categories, dates, matching rules)
crates/storage        storage findings
crates/safety         the ONLY crate that mutates files (validate → approve → journal → execute → verify → undo)
crates/agent-runtime  request engine (intent.rs), model client and catalog
apps/desktop          Tauri app: chat · storage explorer · history · models
native/inference      small llama.cpp worker

🗺️ Roadmap

• Chat that plans, reviews and executes · whole-folder Trash · Put back
• Whole-disk explorer with cached sizes · duplicates · bulk renames · extension changes
• Local models with your own Hugging Face GGUFs
• Signed & notarized builds, auto-update (opt-in)
• Text search inside PDFs/Office docs · near-duplicate photos
• Windows & Linux · “recipes” you can share · keyboard-first mode

More in docs/IDEAS.md — many are marked good first issue.

🤝 Community

Tidy is built in the open and wants your ideas: new request phrases, file categories, model recipes, translations, design feedback, bug reports. Start with CONTRIBUTING.md, open an Idea issue, or say hi in Discussions. If Tidy saved you an afternoon, a ⭐ helps others find it.

🙏 Acknowledgements

Tauri · llama.cpp · Qwen · Hugging Face · lucide · nlohmann/json

📄 License

MIT. Bundled llama.cpp and nlohmann/json licenses ship in apps/desktop/src-tauri/resources/inference/.