dsh-mac-use
Computer-use for DeepSeek Harness on macOS.
Let an agent see the screen, drive the mouse and keyboard, and open apps and URLs —
the same observe → act → verify loop as Claude's computer tool or Codex computer-use,
shipped as a single DeepSeek Harness bundle.
Why
Most computer-use setups ship a separate helper binary that needs its own screen-recording
and accessibility grants. dsh-mac-use doesn't:
- Plugin only, no extra binary. Everything runs inside the DeepSeek Harness Desktop
host process. Screen capture uses the bundled
screencapture tool; mouse/keyboard and
window inspection call macOS APIs through the koffi FFI library (an npm dependency).
- Permissions come from the Harness app itself. Because the plugin runs in the Desktop
host (a child of the Electron app), macOS checks Screen Recording and Accessibility
against DeepSeek Harness. Grant those two once — never to a helper.
- Virtual agent cursor. Mouse actions don't strand your real pointer: clicks land,
then your cursor snaps back, while a floating orange dot glides between action points so
you can watch the agent work — and it disappears when the task is done.
Tools
| Tool | What it does |
|---|
computer_screenshot | Capture the screen (main display, a region, a window, or a display) and return the image plus bounds and pixel→point scale. |
computer_screen_info | Read display geometry (points), cursor position, and the on-screen window list (id, app, title, bounds). |
computer_move | Move the agent cursor to (x, y) in global display points (your real pointer is untouched). |
computer_click | Click at (x, y) — left/right/middle, single or double. |
computer_drag | Drag from (x1, y1) to (x2, y2). |
computer_scroll | Scroll up/down/left/right at a point or the cursor. |
computer_type | Type arbitrary Unicode text into the focused field. |
computer_key | Press a key or chord (letters, digits, arrows, return, tab, escape, F-keys, and cmd/shift/option/control). |
computer_open | Open a URL, file, or app via the system open command. |
A system-prompt section teaches the agent the workflow: observe → locate → convert image
pixels to global points → act → screenshot again to verify.
Requirements
- DeepSeek Harness Desktop (the macOS app), running on macOS (Apple Silicon).
- A vision-capable model for full observe-and-act (e.g.
deepseek-flash, which reads
images). Text-only models can still use computer_screen_info and the action tools.
Permissions (once)
- System Settings → Privacy & Security → Screen Recording → enable DeepSeek Harness.
- System Settings → Privacy & Security → Accessibility → enable DeepSeek Harness
(needed for mouse/keyboard input).
- Fully quit and reopen DeepSeek Harness.
Missing a permission? The relevant tool fails with a message naming the exact setting.
Install
Install through the built-in Plugin Manager — no shell commands, no manual profile edits.
From npm (recommended):
dsh-mac-use
From git:
github:matyik/dsh-mac-use
In the Web sidebar: Plugins → Install bundle, enter either spec above (or a local
directory). The plugin_manager tool's install_bundle accepts the same targets.
The only dependency is koffi@3.1.1 (the version the Desktop runtime ships). It bundles a
prebuilt darwin-arm64 binary and its install script is a fast prebuild check — no
compilation. pnpm 11 may ask you to approve that script once: click Allow these scripts
and retry and it finishes instantly.
Coordinate system
All mouse coordinates are global display points, origin at the top-left of the main
display (the same space as computer_screen_info and the click tools). Screenshots are
returned in device pixels, so computer_screenshot also reports bounds (points) and
scale (image pixels per point):
x = bounds.x + px / scale
y = bounds.y + py / scale
Configuration
| Field | Default | Meaning |
|---|
virtualCursor | true | Use the virtual agent cursor (click, then restore your pointer) and show the floating dot. Set false to leave the pointer where it clicked. |
postDelayMs | 25 | Delay between press/release events. |
scrollAmount | 3 | Scroll wheel units per step. |
typeDelayMs | 0 | Per-character typing delay. |
Security
These tools drive the real desktop — treat them as powerful. They run under the agent's
normal permission/approval policy; keep them in a preset that prompts before acting, and
have the agent re-check its work with a follow-up screenshot. macOS-only: on other
platforms the tools register but error when called.
Contributing
PRs welcome — please read CONTRIBUTING.md first.
License
MIT © matyik
