vision-use

Agents inside DeepSeek Harness have long been trapped in a text-only world: they cannot see the screen, move the mouse, or type into an application, so any desktop-level task such as browsing, filling forms, or writing in Notepad simply cannot be completed end-to-end. vision-use plugs that gap by piping full-screen captures straight into the model's visual channel (real image tokens, not OCR) and handing the agent a full set of desktop tools: smooth cursor glide, left-right-double click, per-key VK injection with zero clipboard writes (Chinese characters go through the IME), and combination-key shortcuts. The result is a genuine see-act-verify loop.

Every hands action force-raises a Codex-style blue glow overlay first, so there is no silent operation. Pressing Esc immediately sets a cancel flag, dismisses the overlay, and rejects all subsequent motion. A task-level keep-alive keeps the overlay visible while the agent is still thinking or calling other tools, then auto-dismisses it about four seconds after the turn ends, with an idle-timer safety net as a backstop.

Designed for users running DeepSeek Harness on Windows native or WSL2 who want the model to independently close out full desktop tasks. The plugin makes zero network requests of its own, has zero runtime dependencies (pure ESM plus a PowerShell child process), and always routes the screenshot to whichever model you have selected-it never picks or alters the routing, keeping privacy fully under your control.