Garden Desk
Open Garden Desk
No account. No key. The model is already there.
Crowded cloud.
Private garden.
Cloud privacy has limits. Your AI coworker stays on your computer.
Free and open source under Apache 2.0. Read every line ↗
No account. No sign-up. No keys.
Private, or simple.
Never both.
So we built both.
The private tools we found were made for developers: install a runtime, pick a model, wire it together. Everything simple sent the work to a cloud. Garden Desk is one download that asks nothing of you. About the name: the cloud is a noisy city with no privacy. Garden Desk is your own desk in a private backyard, where nobody tracks you.
What makes Garden Desk different
-
One download
The app and the model, together.
-
Nothing to configure
Open it. Point it at a folder.
-
Nothing tracked
No analytics. No crash reports. No telemetry.
-
No way online
The model has no route to the internet.
Garden Desk makes no network calls of its own. There is no telemetry, no error report, not even an update check. The agent works inside a virtual machine with no network connection at all, and the folder you select is mounted read-only. The model cannot reach the internet, and it cannot change your originals.
Open it. Point it. Ask it.
That’s the whole onboarding. There is no step four.
Point at your work
Pick a folder or attach files. Your source stays read-only.
Ask, then check the receipts
Describe the outcome. Inspect the answer, the files, every step.
Familiar files, handled locally
- WordDOCX
- ExcelXLSX
- PDFPDF
- ImagesJPG
- DataCSV
Click around the real interface.
The actual desktop app, on invented data. Runs in browser memory: no model, no upload, nothing saved.
Ternary Bonsai 2 runs on your computer.
It runs completely offline. Text, image inspection, and file tasks stay on your computer.
V1 model: Ternary Bonsai 2 27B, an Apache-2.0 open model built from Qwen3.8-27B. Verified on macOS and on Windows.
- Model
- Ternary Bonsai 2 27B
- Context
- Up to 128K tokens
- Operation
- Completely offline
- Setup
- No configuration
Everything else is assembly required.
Seven ways to get the work done. The same nine questions. Checked against public documentation in September 2026.
| Garden Desk | Claude Cowork | Codex for Work | Perplexity Portable Computer | AnythingLLM | LM Studio Bionic | OpenCode + Ollama | |
|---|---|---|---|---|---|---|---|
| What you install | One app | App and account | Web and mobile | App and account | App and server | App and models | Two programs |
| Where the model runs | Local only1 | Cloud only | Cloud only | Hybrid2 | Hybrid2 | Hybrid2 | Local only3 |
| Tuned for one model | Yes4 | Yes | Yes | No | No | No | No |
| Works with no internet | Yes | No | No | Partly5 | Partly5 | Yes | Yes |
| Sealed, no-network sandbox | Yes6 | No | No | No | No | No | No |
| Your files stay read-only | Yes7 | No | No | No | No | No | No |
| Telemetry | None | Yes8 | Yes8 | Yes | On by default9 | Not for local use | No analytics10 |
| Price | Free | $20 to $200 a month | $25+ a seat11 | $20 or $200 a month | Free | Free, or $20+12 | Free |
| Source code | Open source | Closed | Closed | Closed | Open source | Closed, partly open | Open source |
- The model ships inside the download and runs on your computer. There is nothing to pick or install.
- You choose a local or a cloud model, and you install and select it yourself.
- You run a local model server and point the tool at it.
- Prompts, skills, and workflows are tuned end to end for one model. A tool that accepts any model cannot tune for each one.
- The model can run locally, but the app still calls vendor services.
- Agent commands run inside a virtual machine that has no network interface at all. The other tools run commands on your computer or on vendor servers, with a network.
- The folder you select is mounted read-only. The agent writes only to a separate workspace, so it cannot change your originals.
- An account service. Your work runs on the vendor's computers.
- Set
DISABLE_TELEMETRY=trueto turn it off. - Update and model-catalog checks still run until you turn them off.
- ChatGPT Business is $25 for each seat each month, or $20 when you pay yearly. Premium is $125, or $100 yearly. Business needs two seats or more.
- Local models are free. Bionic+ is $20 each month and Pro is $100.
Real privacy costs real silicon.
No datacenter. Your computer does the work. V1 targets:
macOS
Apple silicon Mac
16 GB unified memory or more
Version 1.2.0
Windows
PC with NVIDIA or AMD graphics
16 GB RAM and 12 GB dedicated VRAM or more
Version 1.2.0Do not trust us. Read the code.
Every claim on this page is about software you can open. Apache 2.0, fully public. Go find the telemetry. There is none.