What it does
In plain words.
“How much free disk space do I have?” “Move the installers in Downloads to the trash.” “Turn the volume down to 30.” Arcee carries requests like these out on your own desktop, printing every step as it goes.
- 01
Talk or type
A wake word trained from three samples of your voice, transcripts from Groq Whisper, and spoken replies through Kitten TTS. Or just type: arcee-cli run "…".
- 02
Uses your computer like you do
Files and the shell, your real signed-in Chrome, desktop apps through the accessibility tree, plus mail, calendar, reminders, music and WhatsApp.
- 03
Asks before it touches anything
Reading is automatic. Anything that changes files stops at “Allow? [y/N]”. Commands like rm -rf and sudo are refused outright.
- 04
Any model, any role
By default gpt-oss-20b, qwen3.8-27b and gpt-oss-120b on Groq, rotated within each one's free budget. Pass --core local and everything stays on your machine through Ollama.
How it works
Under the hood.
Not a prompt with tools bolted on: a harness. A planner for long requests, a step loop that calls one tool at a time, and guards around every action.
Every request flows through the same loop, and every step is logged with secrets redacted.
- 01
Hear
sherpa-onnx keyword spotting and Silero VAD wake it; Groq Whisper writes the transcript.
- 02
Plan
Long requests get a plan first; short ones go straight to the loop.
- 03
Act
A step loop calls one tool at a time and reads the result back. Hard steps climb a ladder to a stronger model.
- 04
See
With no accessibility tree to read, a vision model picks cells on an 8×6 grid of the screen, then a zoomed 6×6, to find what to click.
- 05
Guard
Every action passes a policy tier and the shell classifier; web and file content is marked untrusted.
- 06
Remember
SQLite with full-text search. Recall weighs 60% match, 25% recency (14-day half-life) and 15% importance.
- 07
Report
A spoken or written answer, with the full run in arcee-cli history.
Some of the 44 whole-task evals, by name. Each runs in a throwaway home folder.
- trash-installers
- rename-photo
- volume-to-30
- dnd-an-hour
- timer-10-min
- time-in-tokyo
- dollars-to-rupees
- sensex
- drive-to-chandigarh
- news-isro
- injection-in-file
- rm-becomes-trash
Decisions that mattered
Groq by default, local on request
The default core runs on Groq, rotating models by each one's budget (8,000 tokens a minute, 200,000 a day). --core local keeps every request on the laptop.
Context beat a bigger model
A small local model went from 3/8 to 7/8 on the same cases with better context alone, so the harness feeds context before it reaches for size.
Plan only long requests
Planning costs a model call, so it only happens when the request needs one.
Chrome through an extension
A small extension lets Arcee read and act in your real, signed-in browser instead of a fresh headless one.
The rules
What it won’t do.
Three tiers: auto, confirm and never. The never-list holds even when a command hides behind ;, && or $(…).
- sudo
- su
- pkexec
- dd
- fdisk
- mkfs
- shred
- shutdown
- reboot
- passwd
- useradd
- iptables
- ufw
- writes to /etc and /usr
Confirm before changes
Anything that writes, moves or deletes waits for a yes. Deleting goes to the trash, not rm.
Untrusted text can't steer it
If a web page or file contains instruction-like phrases, the next action needs your confirmation, whatever its tier.
Risky clicks ask first
Buttons named delete, send or pay always ask before Arcee clicks them.
Exhibit A
Built in gated phases.
Each phase has a measurable gate it must pass before it counts as done.
- 01Core loop100% valid first replies (108/108) and 15/15 whole tasks on a 1.7B local modelGate passed
- 02Wayland capture and inputLargest miss 1.41 px over 60 clicks; the kill switch stops in 4–6 msGate passed
- 03Browser and desktopFill a web form and rename a file through the GUI, with zero screenshotsIn progress
- 04Daemon, extension, appA fresh install to its first task in under 10 minutesIn progress
- 05Memory and voiceWake to transcript under 1.5 s with his own voice; a week-old fact recalledIn progress
- 06Packaging and release75% on the local evals (it's at 98%) and 20 outside testers for v1.0Next
Still curious?