Turn natural language into real desktop actions.
An AI agent that moves your cursor, types, clicks, sees the screen, and learns from its mistakes — on the Windows apps you actually use.
Thinking process
- Understanding
- Context
- Intent
- Plan
- Considerations
- Execution
The problem
Why AI agents still struggle with real work
Benchmarks like GDPval, OSWorld, WebArena and TheAgentCompany measure exactly this kind of work — and it's still hard.
Brittle grounding
Agents mis-click when layouts shift, scale or theme changes.
Long-horizon collapse
Small errors compound across multi-step tasks until the run derails.
Weak error recovery
Most agents have no memory of what failed, so they repeat it.
Native apps underserved
Browser-first agents ignore the .exe tools real work happens in.
Trust & safety gaps
Opaque actions on a real machine are hard to supervise or stop.
Latency & cost
Heavy vision models on every step make simple actions slow and expensive.
Reproducibility
Without logs and saved skills, a success can't be reliably repeated.
How it works
One loop. Every action.
Every command streams its full thinking process — the same loop you see in the demo videos.
Understanding
Reads what you asked for, in plain language.
Context
Checks the current cursor position and screen region.
Intent
Locks onto the exact target before moving.
Plan
Calculates coordinates and the smoothest path.
Considerations
Applies safe margins so the cursor never sticks.
Execution
Runs the action and reports success with timing.
Features
Built to actually do the work
Intelligent cursor & keyboard control
100+ movement patterns — absolute coordinates, regions, percentages and relative moves, plus typing, hotkeys and clicks.
Real-time reasoning & transparency
Every command gets a visible Thinking Process — understanding, context, intent, plan, considerations — with live success/failure tracking and timing for each action.
Visual perception
OCR, screen analysis, UI element detection and Monet-powered visual reasoning.
Learns from failures
A learning window with Reasoning / Strategy / Progress tabs, retry variations, confidence and mastery tracking. Every success is saved as a replayable skill in SKILL.md.
Microsoft Office automation
Word and Excel flows, formula entry and overlay “thought bubbles” that narrate what the agent is doing.
Try it
Command playground
Type a command or click an example. Runs entirely in your browser.
Slash commands
Press / for everything
34 commands across Cursor, Mouse and Keyboard — e.g. /copy → press ctrl+c.
Agent modes
From simple to visual reasoning
- 1
Simple
Rule-based parsing of everyday commands. Instant and lightweight.
- 2
Enhanced
pynput precision for smoother, more accurate movement and input.
- 3
Reasoning
Streams its full thinking process — understanding, context, intent, plan and considerations — before it acts.
- 4
Visual Perception
OCR and screen analysis to find text and UI elements.
- 5
Real-Time
Thinks before acting and adapts to your patterns, with per-action timing.
- 6
Monet Visual Reasoning
Deeper visual understanding of complex interfaces.
Safety by default
You stay in control.
- Confirmation before risky actions (toggle in Settings)
- Policy gate blocks destructive or sensitive actions
- Native control and coordinate clicks are off unless you enable them
- Press Esc anytime to stop instantly
- Full execution logs for every run
Built for visible, user-directed actions — not unattended high-risk automation.
Editions
Lite or Full
| Lite DEFAULT | Full | |
|---|---|---|
| Agent | Rule-based | Rule-based + heavy AI & vision modes |
| Download size | Small — no multi-GB downloads | Large download |
| GPU required | No | Recommended for vision modes |
| Speed | Fast | Slower first run |
| Best for | What most people want | Experimenting with advanced modes |
Demos
See it work
Live demo — Cursor & Keyboard Agent
Real screen capture of the agent in action: type a command like "move cursor to top left corner" and watch it reason through the request step by step — understanding, plan, and execution — before gliding the cursor exactly where you asked.
Real-Time Reasoning — thinks before it acts
The agent in Real-Time Reasoning mode: for every command it shows its full thinking process — understanding, context, intent, plan, and safety considerations — then executes, while the cursor visibly moves to each corner you ask for.
Interactive Calculator
Natural-language math driven through the Windows Calculator.
Calculator Visual Awareness
Reads the display with OCR to verify each result.
Monet Integration
Visual reasoning over complex, cluttered interfaces.
Enterprise Excel Analyst
Builds formulas and summaries inside Excel workbooks.
Word Writer
Drafts and formats documents directly in Microsoft Word.
Get started
Up and running in three steps
Install the Lite requirements
$ pip install -r requirements-lite.txtRun the agent
$ python run_agent.pyType a command or press “/”
$ move to centerRequirements
- Windows
- Python 3.11+ (3.12 recommended)
- Microsoft Office (optional)
- Tesseract OCR (optional)
The app moves your real mouse and keyboard — keep a hand near Esc the first few times.
Roadmap
What's next
Give your words a cursor.
Turn natural language into real desktop actions.