jiayan-xu/dsh-nuphus-mcp

dsh-nuphus-mcp

dsh-nuphus-mcp bridges dsh to the nuphus MCP server for desktop and browser automation. It registers 36 statically described tools: 15 desktop_* tools and 21 browser_* tools.

Capabilities

Desktop tools cover:

• Screen dimensions and full-screen or regional screenshots.
• Listing, activating, inspecting, moving, resizing, and capturing windows.
• Mouse clicks, movement, hovering, scrolling, and dragging.
• Keyboard input and hotkeys.
• Clipboard writing and clearing.
• Local UI element perception with PaddleOCR and optional YOLO icon detection.
• Optional screenshot analysis through a user-configured OpenAI-compatible vision endpoint.

Browser tools use Chrome CDP and include navigation, interaction, scripted multi-step operations, and accessibility-tree snapshots. Snapshot results assign references such as @1 to interactive elements for later click or input operations, with DOM traversal as a fallback when the accessibility tree is unavailable.

Runtime Configuration

The plugin runs nuphus-mcp.exe as a persistent stdio MCP child process. Requests use newline-delimited JSON-RPC, concurrent calls are matched by request ID, and a later call reconnects after the child process exits.

Set NUPHUS_EXE to override the executable path. The built-in default is:

%USERPROFILE%\.qclaw\workspace\nuphus-mcp\target\debug\nuphus-mcp.exe

Optional vision analysis requires both:

• NUPHUS_MCP_VISION_API_KEY
• NUPHUS_MCP_VISION_MODEL

Plugin logs are appended to ~/.dsh/logs/dsh-nuphus-mcp.log when the path is writable.

Safety

Desktop actions affect the real mouse, keyboard, clipboard, and operating-system windows. Browser automation can execute JavaScript. Operations classified as writes expose a confirm parameter that must be true when nuphus runs in strict-confirm mode.

The dsh bundle registers @jhp830901/dsh-nuphus-mcp with the ID dsh-nuphus-mcp.