Two orders match — it asks which one before writing
“Find Lisa Wong’s order in records.txt and add it to ledger.csv as Date, Customer, Order, Amount, then save.”
Guessing would write the wrong row. Asking costs one sentence.
Open-source computer use for your Mac
Small enough to run on your Mac. Smart enough to ask.
DeskMind reads the screen, works out the next step and acts, all with a small model on your Mac. When a task could mean two things, it asks you instead of guessing.
What makes it different
01
A small model, on your Mac
A 0.8B model decides each step and a 4B checks the hard ones. Both run on your Mac: no cloud round-trip, no per-step bill.
0.8B decides in ~0.5 s · 4B checks in ~3.6 s
02
System One: choices, not guesses
Each step is a multiple-choice question: the model scores every option instead of writing text, so each one gets a probability. Unsure steps go to the 4B or to you, and any agent can call it through /v1/systemone.
No generated text · 0 false “done” in 39 real-desktop runs
03
Open, from eyes to hands
Eyes, Brain, Hands and the Mac app are all open source, along with Bench, which grades them. Every result lists its sample size, so you can reproduce it.
5 repositories · Apache-2.0 code
How it works
One loop, three parts, packaged in one app. A separate bench measures the whole loop on a real desktop.
EyesSees the screen. Finds the target when there is no accessibility tree.BrainDecides the next step: a probability for every option, a question when unsure.HandsActs on your desktop and checks the result before calling it done.
AppAll of it, in one Mac app: permissions, models and the run loop, ready to use.Three real runs
Recorded with the released model. Waits are shortened in the videos; nothing else is cut.
Find a song’s live version and play it
“Open NetEase Cloud Music, search Billie Eilish’s BIRDS OF A FEATHER and play the live version.”
No API, only pixels: it reads the screen and picks the live track from look-alike results.
Copy a web table into a CSV file, then save and close it
“Open parts.csv from the attached folder, append the table’s four rows sorted by Qty from high to low, then save and close it.”
Across two apps: read in Safari, open the file from a folder, write, save, close.
Results, with the sample size
Measured on one Mac (M4 Pro, 48 GB) with the released 0.8B → 4B router. Small samples; we say where it still fails.
39/39
real-desktop runs passed, strict graders (bench v25, 13 tasks × 3)
0
times it said “done” when the task was not done (same 39 runs)
0.835
JevBench public items, 4B model (231 items; no sealed score yet)
Still hard
- Copying long tables (more than about four rows) or filtered rows
- Filling a form from a photographed receipt
- Steps the 4B has to check take a few seconds
How we measure
Graders check the final state of files and apps. Every attempt counts, environment failures are listed, and the bench is open.
All results and methods →Your Mac does the thinking
Inference runs locally by default. Here is exactly when anything leaves your machine.
- Models download onceFrom Hugging Face, or ModelScope in mainland China. After that, deciding a step needs no network.
- A cloud model is opt-inThe app never calls one. If you point the router’s escalation tier at a cloud model yourself, the steps routed to it go to that service.
- Apps behave as they always doA web page or music app it drives still talks to its own servers.
See. Think. Act.
得心，应手。 Download the app, or start with the models and the code.