Nuphus Desktop Automation MCP

What does this agent do, and when should you use it?

nuphus-mcp is a cross-platform stdio MCP server that exposes desktop control and Chrome automation through 38 standard MCP tools. Its Rust workspace contains the MCP server, the CDP-based nuphus-browser component, and the desktop-api control layer. It exchanges single-line JSON-RPC 2.0 messages over stdin and stdout, with no daemon or HTTP service, and is documented for clients including Claude Desktop, Cursor, VS Code, and Copilot. Desktop and browser operations require no API key; desktop_perceive runs PaddleOCR locally with optional YOLO icon detection, while desktop_vision connects to a user-supplied OpenAI-compatible or Anthropic-native vision model. Browser and desktop support is complete on Windows, macOS desktop input requires Accessibility permission, and Linux window and input capabilities are limited.

An MCP client sends JSON-RPC requests such as initialize, tools/list, and tools/call to nuphus-mcp. Its 15 desktop tools inspect screen size, produce PNG/base64 screenshots, enumerate and manipulate windows, and perform mouse clicks, drags, scrolling, keyboard input, hotkeys, and clipboard operations. desktop_perceive runs local PaddleOCR to return text elements with coordinates and can load a YOLO icon model through ONNX Runtime; desktop_vision sends a screenshot to the configured vision model for semantic interpretation. The 23 browser_* tools use nuphus-browser and chromiumoxide CDP to launch or attach to Chrome or Edge, navigate pages, create accessibility snapshots with @N references, click, type, execute scripts, extract content, and manage tabs, cookies, uploads, and downloads. The process reads newline-delimited JSON from stdin, writes JSON-RPC responses to stdout, sends logs to stderr, and validates paths used for screenshots, uploads, and file drags.

1. A Claude Desktop, Cursor, or other MCP user who wants a model to inspect the screen and operate windows, the mouse, and the keyboard.
2. An automation developer who needs DOM-level browser workflows for navigation, form entry, extraction, cookies, tabs, uploads, and downloads in managed Chrome.
3. A user who needs to control an existing browser session through an external CDP endpoint, or fall back to OCR-driven visual control when CDP is unavailable.
4. An agent developer who combines desktop_vision with desktop_perceive to obtain both semantic understanding and pixel-precise action coordinates.
5. A DeepSeek Harness user who installs the dedicated dsh-nuphus-mcp plugin or mounts this server manually through its stdio MCP client.

How do you install or deploy this agent?

The recommended installation uses npm to select a prebuilt platform binary and does not require a Rust toolchain:

npm install -g @nuphus/nuphus-mcp

Start the stdio server after installation:

nuphus-mcp

A source build is also supported and requires the stable Rust toolchain:

cargo build --release -p nuphus-mcp

Chrome or Edge is required for browser tools. desktop_vision needs NUPHUS_MCP_VISION_API_KEY and NUPHUS_MCP_VISION_MODEL only when that tool is called. On its first run, desktop_perceive downloads the OCR models and icon_detect.onnx.

How do you use this agent?

Start with strict confirmation enabled so write operations must include confirm: true:

nuphus-mcp --confirm-write

A Claude Desktop configuration can launch it as follows:

{
  "mcpServers": {
    "nuphus-mcp": {
      "command": "nuphus-mcp",
      "args": ["--confirm-write"]
    }
  }
}

Use the documented initialization request for a quick smoke test:

echo '{"jsonrpc":"2.0","id":0,"method":"initialize","params":{"protocolVersion":"2024-11-05","capabilities":{},"clientInfo":{"name":"test"}}}' | nuphus-mcp

To call desktop_vision, provide a vision API key and model ID in the MCP process environment:

export NUPHUS_MCP_VISION_API_KEY="your-api-key"
export NUPHUS_MCP_VISION_MODEL="your-vision-model"

The default path uses an OpenAI-compatible Chat Completions endpoint at https://api.openai.com/v1. For Anthropic's native Messages API, set NUPHUS_MCP_VISION_BASE_URL to https://api.anthropic.com/v1; provider detection can use the hostname, or NUPHUS_MCP_VISION_PROVIDER can be set explicitly to anthropic.

What are this agent's strengths and limitations?

How does this agent compare with similar options?

Key facts side by side with the most closely related agents.

How does FollowAgents rate this agent?

Recent releases

Latest entries from the project's GitHub Releases. Notes are quoted from upstream as written, not rewritten or tested by FollowAgents.

1. v0.3.3 2026-10-13
   Release notes excerpt

   Full Changelog: https://github.com/mrpulor-gh/nuphus-mcp/compare/v0.3.2...v0.3.3

   View on GitHub ↗
2. v0.3.2 2026-10-11
   Release notes excerpt

   Full Changelog: https://github.com/mrpulor-gh/nuphus-mcp/compare/v0.3.1...v0.3.2

   View on GitHub ↗
3. v0.3.1 2026-09-28
   Release notes excerpt

   What's Changed * fix(browser): serve on a big-stack thread and bound the remaining waits — fixes #10 by @mrpulor-gh in https://github.com/mrpulor-gh/nuphus-mcp/pull/11 New Contributors * @mrpulor-gh made their first contribution in https://github.com/mrpulor-gh/nuphus-mcp/pull/11 Full Changelog: https://github.com/mrpulor-gh/nuphus-mcp/compare/v0.3.0...v0.3.1

   View on GitHub ↗

FAQ