The Ultimate Front-End OS Engine for AI Agents. 0 Vision Tokens. 733× Faster. 100% Deterministic.
FEOM is a production-grade front-end operation architecture for Windows AI Agents. This repository contains the Open-Core Edition, allowing developers to test and verify real-time structural OS routing (achieving ~8ms background UIA invocation), completely eliminating VLM vision token costs.
| Feature | 🟡 Open-Core (This Repo) | 🟢 FEOM Pro / Enterprise |
|---|---|---|
| MCP Server Framework | ||
| ✅ Included | ✅ Included | |
| UIA Tree Structural Read | ||
| ✅ Included (~75ms) | 🚀 Predictive Async Cache (<1ms) | |
| Win32 SendInput Click | ||
| 🟢 Standard (~3.96ms) | ⚡ Zero-Alloc UltraClick (<1.0ms) | |
| Background UIA Invoke | ||
| 🟢 Basic (~8ms) | ⚡ Win32 PostMessage (~1ms) | |
| Multi-Monitor High DPI | ||
| ❌ Static Coords Only | ✅ Dynamic Scaling Matrix | |
| Taskbar Push Auto-Recovery | ||
| ❌ Fixed Offsets | ✅ Icon Vector Tracking | |
| Input Pipeline Optimization | ||
| ❌ Raw Character Streams | ✅ Clipboard Parallel (-71%) |
Always show, then report. Every sequence must execute visibly in the foreground.
Never click blindly. Structural state verification is forced prior to any input.
Alt+F4 blindly kills whatever window has focus — frequently closing the orchestration engine itself.
| Execution Stage | Traditional MCP Agent | FEOM Hybrid Mode | Time Saved | |---|---|---|---| | App Bootup | ~3.0s (UI traversal) | 0.01s | 99.6% | | Form Input | ~5.0s (Vision Loop) | 1.2s | 76.0% | | Close & Verify | ~4.0s (Screenshot) | 2.0s | 50.0% | Total End-to-End | ~12.0s | ~3.21s | 73.2% Faster |
terminal("start calc")
| Requirement | Traditional VLM | FEOM Open-Core |
|---|---|---|
| GPU | RTX 3060+ required | None - any CPU works |
| RAM | 16GB+ | 4GB |
| Network | Required (API per click) | Zero - fully local |
| Cost/op | $0.01-$0.05 (vision tokens) | $0 |
| Display | Yes (needs screenshot) | No - works headless |
| Hardware cost | $1,000+ gaming PC | $50 used Dell |
| Annual license | $15K+ (RPA) | $0 (Open-Core) |
| VLM Screenshot | FEOM OS-Native | |
|---|---|---|
| Output | Coordinate guessing | AutomationID targeting |
| Repeatability | Different each time | 100% identical |
| DPI scaling | Coordinates drift | UIA auto-maps |
| Background | Cannot work | Works minimized |
| Production-ready | Demo only | Production grade |
| Tier | Product | Price |
|---|---|---|
| 1 | Turbo-MCP Plugin (anti-crash + DPI matrix) | $14.99/mo |
| 2 | AI-SaaS Task Force (legacy ERP/CRM migration) | Custom |
| 3 | Enterprise Consulting (stability + multi-monitor audit) | $3K-$10K |
| 4 | Hermes-Split-OS SDK (parallel multi-agent framework) | License |
Inquiries: Open B2B Ticket or a92070888@gmail.com
pip install pywinauto pywin32 mcp
python server.py
Then add to MCP client config:
{
"mcpServers": {
"feom": {
"command": "python",
"args": ["path/to/server.py"]
}
}
}
No GPU. No cloud API. No Docker. Any Windows PC.
| Tool | Description | Latency |
|---|---|---|
| feom_click | Click at screen coords via SendInput | ~4ms |
| feom_uia_invoke | Invoke UIA element by AutomationID | ~8ms |
| feom_launch | Launch app via terminal pipe | ~0.01s |
| feom_list_windows | List all visible windows | ~75ms |
pip install pywinauto pywin32 mcp python server.py