Ep1cGaimer

cu-s1 is the fast local hand for a planning agent. Today's computer-use agents send every screenshot to a frontier model and wait for it to choose each click: slow, expensive, and a waste of the planner's attention. cu-s1 splits the job. A frontier planner (I used Muse Wonder) writes a short chain of steps; cu-s1 executes it on a Windows PC with a 2B vision model (Qwen3-VL 2B via Ollama) running locally: screenshot in, native mouse and keyboard out, each step checked on screen. If the screen doesn't match, a target can't be confirmed, focus or window geometry changes, or a step is flagged risky, it stops and hands back to the planner with a reason and a screenshot instead of guessing. It never replays an ambiguous action and never treats a pixel change as proof of success. It ships as an MCP server (windows, look, run_steps, stop) and a local control panel that share one execution engine, with live per-step progress, a monitor view for plans arriving over MCP, cooperative stop and MCP cancellation. No accessibility APIs or app integrations: pixels only, so it works across Windows apps. Measured (raw records in bench/runs): a live Wikipedia task 3/3 in 9.5 s; Windows Calculator 9/9 visual clicks with the answer verified on screen (16.2 s); File Explorer 3/3 (6.0 s); a request to click a non-existent button was handed back with no input sent. On matched supplied-plan holdout tasks, 3/3 at a 6.8 s median versus 25.6 s for built-in computer use; reserved holdout 8/9. These are small benchmarks, and the repo states the limits: Windows only, the 2B model is fallible, drag and multi-monitor not yet verified live.