How Neverclick Rethinks AI Consumer Hardware by Ditching Accessibility APIs
Lazo Velko didn't build Neverclick to chase software trends. He built it because his wrist hurt. Dealing with a severe repetitive strain injury that made pushing a mouse unbearable, the developer built a utility that lets users perform mouse actions across an entire Windows display using nothing but keyboard shortcuts.
The utility is completely free, works offline, serves zero ads, requires no account signup, and collects no user data. According to The Register's report, the GitHub repository currently exists solely for issue reporting rather than hosting full source code. That might frustrate open-source purists, but Velko's personal pitch is simple: "I recovered years ago, and I owe it to this software. I still use Neverclick every day and can't imagine using my computer without it."
What makes the project stand out isn't just its origin story—it's the architecture. Instead of hooking into standard operating system accessibility pipelines, Neverclick runs a custom computer vision model that analyzes raw pixels directly on your display.
Why AI Consumer Hardware Developers Are Moving Beyond Windows Accessibility APIs
Most keyboard navigation utilities rely on Windows UI Automation or traditional OS accessibility APIs to locate buttons, menus, and input fields. Velko tried that route and abandoned it quickly. In a Hacker News discussion, he called accessibility APIs clunky, slow, and unpredictable, noting that direct computer vision eliminates those structural dependencies entirely.
Velko built the vision engine himself without third-party libraries. His central claim: raw pixel analysis on screen elements actually outperforms legacy accessibility frameworks on older machines. Users running Neverclick on decade-old PCs report that the computer vision model triggers instantly, whereas Windows UI Automation frequently stutters or lags.
It's a stark contrast to web-based AI systems. While the accessibility tree decides whether an AI agent can read your page by providing structured ARIA roles for web scrapers, Neverclick takes the exact opposite approach on the local desktop: skip the OS accessibility tree and read pixels directly off the glass.
The app footprint remains remarkably light. The download requires about 40 MB of disk space and uses roughly 200 MB of RAM on a 1080p display. Because it doesn't run in a continuous loop, CPU utilization stays near zero until you press an activation shortcut.
Hints, Shortcuts, and the Pixel Overlay Mechanics
Activating Neverclick relies on explicit key combinations. Pressing Ctrl+Enter triggers a left-click overlay, while Alt+Enter executes a Ctrl+Left Click. The app also supports shortcuts for switching windows and selecting multiple points on screen to close several windows simultaneously. There is no right-click mapped by default, though users can configure custom bindings.
When triggered, the app covers screen elements with two-letter "hints" (such as ek, af, or jj). Typing the matching letter pair moves the cursor directly to that target—whether it's a button, menu item, or word. By default, word hints land in the center of a word, but pressing a trailing number key targets individual characters. Users can also restrict hint overlays to the active window or let them span all connected displays.
During testing by El Reg US Editor Avram Piltch, the system handled targeted clicks cleanly across standard Windows apps. However, several limitations remain. Drag-and-drop operations and text highlighting are not supported yet, though Velko has mapped text selection—choosing two hints to highlight text in an active editor—for future updates.
Architectural Tradeoffs and Visual Overcrowding
Neverclick isn't without friction. The visual overlay can be jarring. When two-letter hint codes cover every interactive element, reading background text or identifying small icons becomes difficult. To clear the hints, you must execute a click or exit the mode.
Shortcut collisions present another hurdle. The default Ctrl+Enter binding overrides standard app shortcuts. In Gmail, for example, Ctrl+Enter normally sends an email; with Neverclick running, it pops up the hint overlay instead. Rebinding shortcuts resolves the conflict, but requires manual tweaking.
Security-conscious users will also note the closed-source nature of the download. While Velko communicates openly on forums, trusting a compiled binary that inspects screen pixels requires baseline confidence in the developer. For everyday task management, that tradeoff may be acceptable, but enterprise environments may hesitate before deploying non-auditable desktop binaries.
Desktop Vision Models in the Ecosystem
Neverclick's pixel-first approach mirrors a broader shift across independent research teams and agentic labs. As models like Google Gemini, Anthropic's Claude, and OpenAI's ChatGPT or GPT-5 evolve toward multimodal screen understanding, software interaction is shifting from API wrappers to direct vision analysis.
Consider how developers work today. When writing research notes, you might draft text in an editor, polish phrasing using QuillBot, verify originality with a Free AI Detector, or run a final pass through a structural Checker. Jumping between those tools manually wastes focus. Physical input devices like Project Mirage's Dune keypad address this by providing dedicated macro keys for macOS users. Neverclick provides a pure software alternative for Windows, removing mouse dependency without requiring external hardware.
Educational resources from platforms like DeepLearning.AI have democratized local vision pipelines, enabling solo developers to construct specialized models from scratch. Interestingly, Velko describes himself as a "self-professed C++ hater," making his choice to build a high-performance C++ desktop utility a quirky bit of practical engineering.
As automated workflows scale—such as when Abridge AI streamlines medical documentation, or when Cheq launches agent intent tools as part of a comprehensive identity intelligence and security solution for the agentic web—giving human users efficient physical or keyboard control remains critical. Automated agents need clear boundaries, and lightweight desktop utilities ensure humans can navigate their own displays at speed without relying on clunky OS accessibility frameworks.
What's Next for Neverclick
Velko is currently developing a macOS version of Neverclick, though he hasn't committed to a firm release date. He notes that he is currently overworked while completing core features for the Windows build, including window dragging and dual-hint text selection.
Once the macOS version launches, users could theoretically combine Neverclick with head-tracking utilities—such as recently highlighted AirPods tools that let users scroll screens by tilting their heads up and down—to build a completely hands-free desktop workflow.
For now, Neverclick remains a specialized Windows utility. It won't replace a mouse for gaming or precision graphic design. But for users suffering from wrist strain, or for keyboard enthusiasts looking to streamline their daily desktop workflow, Velko's 40 MB experiment offers a fast, offline, and surprisingly practical alternative to legacy accessibility APIs.
