AI Agent Hub
Back to plugins
🧩

dsh-plugin-computer-use

admin-security Updated 2026.09.14

Run the following command in DeepSeek Harness:

dsh plugin install aibo204/dsh-plugin-computer-use

Paste the following prompt into your AI chat to install this plugin:

Run dsh plugin install aibo204/dsh-plugin-computer-use inside your DeepSeek Harness project to install this plugin; the source code is available at https://github.com/aibo204/dsh-plugin-computer-use

About this plugin

dsh-plugin-computer-use gives DeepSeek Harness agents an opt-in path to the user's live desktop. Once explicitly authorized within an Agent Preset, the agent can observe running application windows, read the accessibility element tree, and perform input actions such as clicks, keystrokes, scrolling, and dragging, extending conversations beyond the chat window into the real desktop environment.

Under the hood the plugin launches the Open Computer Use native MCP runtime and bridges nine stable tools under the mcp__computer_use__ namespace, covering state observation (list_apps, get_app_state) and control actions (click, type_text, press_key, scroll, drag, type_text, set_value, perform_secondary_action). Safety is layered: accessPolicy offers per-call, observation-only, or full-allow approval modes; highRiskConfirmation enforces a mandatory second check for send, delete, purchase, and upload controls; allowedApps and deniedApps filter exact target strings before any approval prompt is shown. Every action produces an audit summary, optional automatic screenshot evidence, and a private local archive, with sensitive payloads from type_text and set_value redacted from approval reasons and trajectory views.

It is best suited for developers and power users who already author Agent Presets in DeepSeek Harness and want their models to safely assist with desktop-level tasks, such as driving a mail client, file manager, or IDE under tight scope. The plugin binds a single MCP process and snapshot namespace to one Preset instance, so desktop tools never leak to the Host root and remain confined to the sandbox you have intentionally defined for a given session.

Use Cases

  • Authorize an agent to drive a local mail client through a sending workflow under strict approval scope
  • Let a model read an IDE window's accessibility state and perform safe code edits and build actions
  • Run desktop-level automation tasks across macOS, Linux, and Windows

Best For

  • Developers already authoring Agent Presets in DeepSeek Harness
  • Power users who need models to safely operate local applications within a sandbox
  • Engineering teams building controlled desktop automation pipelines