mrpulor-gh/nuphus-mcp

Desktop automation MCP server — computer use for any AI agent: control screen, windows, mouse/keyboard, and Chrome via Model Context Protocol (stdio). Not a DSH plugin — DSH users install dsh-nuphus-mcp.

nuphus‑mcp – Desktop & Browser Automation Server for AI Agents

What it is – nuphus-mcp is a lightweight, cross‑platform server that lets any MCP‑compatible AI agent (Claude Desktop, Cursor, VS Code Copilot, etc.) control the local computer. It talks JSON‑RPC 2.0 over standard input/output, so there is no network daemon to run – just a single binary.

Why it matters – Modern LLM‑based agents need to see the screen and act on it (move the mouse, type, click, read OCR, drive Chrome). nuphus-mcp provides that bridge while keeping the heavy lifting (vision, OCR) local, so no external API keys are required for desktop work. Vision can be hooked to any OpenAI‑compatible or Anthropic model you supply (BYOK – bring‑your‑own‑key).

Core Features

Category Tools (total) Highlights
Desktop 15 Screen size, PNG/base64 screenshots, window list & manipulation, mouse click/drag/scroll, keyboard input, hotkeys, clipboard write/clear. Implemented via the desktop-api Rust crate (xcap + Win32, no Tauri).
Vision – desktop_vision forwards a screenshot to your own vision model (OpenAI‑compatible or Anthropic). desktop_perceive runs PaddleOCR (and optional YOLO icon detection) locally, auto‑downloading models on first use.
Browser 23 Full Chrome/Edge CDP control via nuphus-browser (chromiumoxide): navigation, DOM‑level clicks, typing, script execution, scrolling, accessibility‑tree snapshots, cookie management, tab handling, file uploads/downloads, etc.
Transport – JSON‑RPC 2.0 over stdin/stdout – no HTTP server, no background daemon.
Safety – Destructive tools are annotated; optional --confirm-write (or NUPHUS_MCP_CONFIRM_WRITE=1) forces an explicit `

Related

  • Project
    ASLP-lab/MeanVC

    MeanVC is a lightweight, zero-shot voice conversion system that uses mean flows and a diffusion transformer to enable real-time, single-step timbre transfer.

  • Project
    ucsandman/DashClaw

    DashClaw is a self‑hostable guard that intercepts potentially dangerous tool calls from AI coding agents (Claude Code, Codex, Hermes, OpenClaw, etc.). It blocks or queues the action, routes it to a remote approvals inbox, and records a signed decision ledger. The project provides Node/Python SDKs, REST API, and plugins for major agent runtimes, enabling safe unattended agent runs with audit‑ready evidence.

  • Project
    useagenthq/useagent

    useAgent is an open‑source, self‑hosted AI coworker platform that lets LLM agents (Claude Code, Codex, OpenCode, Pi) operate inside isolated Linux sandboxes, browse the web, run commands, edit code, and interact with team tools (Slack, GitHub, Gmail, Notion, etc.). It provides a UI with a live timeline, terminal, and workspace, stores every run in PostgreSQL for replay, and supports reusable “skills” and a shared knowledge base. The project is written for the Bun runtime, uses pgvector for embeddings, and is released under AGPL‑3.0 with an optional commercial license.

  • Project
    gendigitalinc/sage

    Sage is a security plug‑in for AI coding assistants (Claude Code, Cursor, VS Code, OpenClaw, OpenCode). It intercepts tool calls—shell commands, URL fetches, file writes, package installs—and runs multiple threat checks (cloud URL reputation, 300+ local YAML heuristics, a prompt‑injection ML model, supply‑chain analysis, and Windows AMSI). If a request looks unsafe, Sage blocks it and logs the decision. Installation is done via the respective assistant’s marketplace or npm, and the project is open‑source under Apache 2.0.