What can an AI agent that controls your whole computer actually do?

What the thing actually is

Most AI tools live inside a browser tab. This one lives on your desktop. It captures the screen in real time, builds a UI tree of what it sees, and then drives the mouse and keyboard through a native HID layer. The result is a perceive-decide-execute-verify loop that closes without a human in the middle. The project calls itself a digital operator rather than a coding assistant or chatbot.

The current release carries version v26.9.30-1 on Windows and v26.10.1 on macOS. Both builds dropped on 30 September 2026. The Windows installer weighs 374 MB. The macOS image weighs 849 MB. SHA256 checksums are published for each file so you can verify the download before running it.

How the toolset breaks down

The capability matrix lists twelve domains.

PC System Control contributes 57 tools covering processes, services, registry, windows, network, WiFi, disk, task scheduler, software management, archives, and notifications. Browser Automation adds another 57 tools built on the full Playwright suite: page operations, cookie and storage handling, network interception, JavaScript execution, PDF export, device emulation, and geolocation.

Office Suite brings 23 tools for Word, Excel, PowerPoint, and PDF with full create-read-update-delete, charts, form filling, template replacement, and PDF merge-split-watermark. Media Processing supplies 10 tools for transcode, trim, merge, extract, thumbnail, volume, and subtitle extraction. AI Generation covers six tools across text-to-image via FLUX, Stable Diffusion, DALL-E, and Midjourney, plus background removal, resize, watermark, TTS synthesis, and quality check.

Video Generation adds four tools: Seedance 2.0 text-to-video, image-to-video, video enhance, and subtitle extraction. Screen Perception provides 12 tools for real-time capture, UI tree, focused element, element finder, continuous monitor, and AI visual analysis. HID Control delivers 15 tools for mouse move, click, right-click, double-click, drag, scroll, keyboard type, combo, press, app launch, and clipboard input.

Memory and Knowledge contributes 10 tools spanning L1 short-term with full-text, vector, and hybrid RRF retrieval, L3 persistent knowledge base, and PDF-Word-Excel ingestion with auto-vectorization. LSP Code Intelligence adds five tools for workspace symbol search, diagnostics, find references, go-to-definition, and hover type info. Core Utilities rounds out the list with 18 tools for file read-write-search-replace-diff, command execution, clipboard, and AST-tsc-cargo validation. The Skill System itself contributes two tools: a skill store with 30-plus official skills and an execution engine.

What the architecture looks like

The core binary is written in Rust and named moching.exe. Inside it sit an IPC engine, a HID input subsystem, and a screen capture subsystem. These talk through a native IPC bridge to an embedded Python runtime. That runtime hosts three components: an agent engine, the 219 tools, and a knowledge store. Rust handles speed-critical paths. Python handles flexibility and the tool implementations.

The design is described as 100 percent independently developed. It is not a wrapper around another agent and not a modification of an existing codebase. The project emphasizes that the moat is not any single capability but the complete loop: see screen, understand, take action, verify result.

What you need to run it

System requirements are explicit. Windows 10 or later on x64. macOS 12 or later on Apple Silicon. Four gigabytes of RAM. Any OpenAI-compatible API key — that includes OpenAI itself, Anthropic Claude, Google Gemini, or any endpoint that speaks the same protocol. The installer walks you through download, launch, key entry, and then you type a request and watch it work.

A 14-day free period starts on sign-up. After that the model card does not state a price; it only says bring your own model key. The referral program offers 500,000 credits to both referrer and referee after the referee completes a first top-up. Credits are described as gifted rewards, not cash, and do not count toward the top-up threshold.

What the license and pricing mean

The license field reads proprietary. All rights reserved. The model card does not publish source code. It does not publish a privacy policy link beyond the generic website footer. It does not disclose telemetry details. It does not disclose whether screen captures or keystrokes ever leave the machine. It does not disclose how many installations exist or how many active users there are.

Support channels are an email address noreply@mochingcode.com, GitHub Issues, and the website mochingcode.com. The changelog shows two entries: the current release and a previous one from 15 September 2026 labeled v26.9.16-1 with the same 219-tool count. No earlier history is shown. The project asks users to star the GitHub repository for visibility and to like the model card on Hugging Face to help it trend.