Minus Alex

Minus Alex grew out of a much larger project called WhisperShell. The original idea was simple: instead of forcing hardware to adapt to software, make the software adapt to whatever hardware is already available. I wanted to see how much computer control I could give to a very small local AI without requiring powerful hardware, cloud dependence, or a huge software stack. Minus Alex became the stripped-down contest version of that idea: a local Resident that can observe, choose actions, and pass those actions through First Touch. Minus Alex connects a small local AI Resident to an MCP-based control layer and First Touch. The Resident can propose actions such as changing volume, brightness, Bluetooth state, or interacting with mapped screen controls. First Touch remains the execution boundary. The AI can suggest what it wants to do, but First Touch determines what actually happens. The demo also includes voice control for actions such as opening a tab, selecting a floating tab, play, stop, minimize, and close. The project is deliberately lightweight and modular. Some heavier WhisperShell components were replaced with simpler alternatives for this build so the system could run reliably on modest hardware. The system was built bootleg-style on existing Windows hardware using local tools and a small Ollama model. The main pieces are a local AI Resident using Qwen, an MCP bridge for tool discovery and requests, First Touch for controlled execution, screen mapping for predefined visual controls, local voice commands, Windows-based system controls, and a lightweight web interface for testing and demonstration. The architecture is modular, so individual capabilities can be enabled, replaced, or removed depending on the hardware available. The goal was not to create another large assistant stack. It was to prove that a small local AI can be given useful perception and action paths without requiring expensive hardware. The biggest challenge was working with limited hardware. Running the AI, interface, screen-control components, recording tools, and video capture at the same time was too heavy for the machine, so I had to find lighter routes. That meant simplifying some parts, using screenshots where necessary, filming some interactions externally, and choosing lightweight alternatives instead of loading every possible component. Another challenge was screen-coordinate mapping across multiple monitors and different scaling behavior. First Touch had to work with the reality of the machine rather than assuming a perfect environment. Those limitations became part of the design philosophy: use the cheapest signal and smallest component that still solves the problem. I learned that local AI does not have to be large to be useful. A small model becomes much more capable when it is connected to carefully scoped tools, perception, and execution. I also learned that modularity matters more than raw size. A system that can turn capabilities on only when they are needed can run on hardware that would otherwise be considered too weak for this kind of experiment. Minus Alex is only one small branch of WhisperShell. The larger architecture includes First Sight, First Hearing, First Touch, local Residents, shared context, robotics, and additional device adapters. The next step is to rebuild WhisperShell around the same lightweight principle: everything available, nothing loaded until it is actually needed. The long-term goal is simple: Give AI something to understand, and safe ways to act on screen using hardware people already own.

Built With

• alexa+
• javascript
• mcp
• node.js
• ollama
• python
• qwen
• windows