mrpulor-gh/nuphus-mcp

Desktop automation MCP server — computer use for any AI agent. See the screen, control windows/mouse/keyboard, and drive Chrome over the Model Context Protocol (stdio). Desktop & browser automation need no API key; OCR runs locally; vision plugs into your own vision LLM (OpenAI-compatible or Anthropic native, BYOK).

nuphus-mcp is a lightweight, cross-platform desktop automation MCP server that exposes desktop + browser automation as standard MCP tools. It speaks JSON-RPC 2.0 over stdio — no daemon, no network service, one binary. Claude Desktop, Cursor, VS Code, Copilot, or any MCP client can connect and immediately control the screen, windows, keyboard/mouse, and Chrome — computer use for any AI agent — desktop & browser automation need no API key; local OCR is built in; vision works with your own vision LLM (OpenAI-compatible, BYOK).

🇨🇳 Mainland China mirror: this repo is mirrored on Gitee for fast in-China access (Chinese docs served by default there). 中文文档

Part of the Nuphus ecosystem: Nuphus — a local-first AI agent with real desktop execution and dual-device (phone-as-second-screen) sync. nuphus-mcp exposes the same desktop + browser automation to any MCP client.

Using DeepSeek Harness (DSH)? nuphus-mcp itself is a plain stdio MCP server — not a DSH/Cordis plugin. Install the dedicated dsh-nuphus-mcp plugin instead (see DeepSeek Harness (DSH)).

┌──────────────────┐   stdio JSON-RPC   ┌──────────────────────┐
│  Any MCP Client  │  ───────────────►  │      nuphus-mcp      │
│  (Claude/Cursor/ │  ◄───────────────  │  desktop-api crate   │──► screen/window/mouse/keyboard
│   Nuphus itself) │  single-line JSON  │  nuphus-browser crate│──► Chrome (CDP)
└──────────────────┘                    └──────────────────────┘

• 45 MCP tools (22 desktop + 23 browser) — screenshots, window control, mouse/keyboard, Chrome CDP automation, and more — see TOOLS.md / TOOLS.zh-CN.md for the full reference.
• Desktop automation: screen size, screenshot (PNG/base64), window list, window activate/screenshot/move/resize/info, mouse click/drag/scroll/position, keyboard input/hotkey, clipboard write/clean — implemented on the desktop-api crate (xcap + Win32, no Tauri dependency).
• Semantic desktop (Accessibility/UIA) — 7 tools that drive native UI through the platform accessibility tree instead of pixel coordinates: desktop_targets_list, desktop_target_bind, desktop_semantic_observe, desktop_semantic_candidate, desktop_semantic_execute, desktop_semantic_action, desktop_verify_state. One contract — observe → pick a candidate → execute → verify — where native window handles, coordinates and platform locators never cross the MCP boundary; the caller only sees opaque ids and stable semantic locators. Backed by Windows UIA and the macOS Accessibility (AX) API.
• Computer vision pair: desktop_vision (BYOK — send a screenshot to your own vision model via an OpenAI-compatible or Anthropic native API) + desktop_perceive (local OCR with PaddleOCR, models auto-downloaded on first run; optional YOLO icon detection). Used together they give AI agents both semantic understanding and pixel-precise coordinates — the battle-tested vision→perceive flow from the Nuphus desktop app. See TOOLS.md for BYOK env vars, model setup, and the recommended flow.
• Browser automation: navigate, snapshot (accessibility tree with @N refs), click, type, exec, scroll, extract, screenshot, evaluate, back/forward, wait_for, cookies get/set/import, upload, tabs, downloads — implemented on nuphus-browser (chromiumoxide CDP).
• Zero-cost stdio: no HTTP server, no daemon. The process reads single-line JSON from stdin and writes responses to stdout.
• Safety-first: destructive tools are annotated per the MCP spec; optional strict-confirm mode; path validation for screenshots, uploads, and file drags.

nuphus-mcp/
├── Cargo.toml                  # workspace root
├── TOOLS.md / TOOLS.zh-CN.md   # 45-tool reference
├── crates/
│   ├── nuphus-mcp/             # MCP Server (this repo's product)
│   ├── nuphus-browser/         # Browser automation core (CDP)
│   └── desktop-api/            # Desktop control core (vendored)
└── ...

• Rust toolchain (stable) — build from source with Cargo.
• Chrome or Edge — required for browser tools. The server auto-detects an installed browser; if none is found, browser_* tools return a clear error.
• Windows recommended for full desktop control — see Platform Support below.

Execution HUD: non-intrusive execution feedback on every platform — Windows shows a compact status card anchored to the bottom-right of the work area, clear of the taskbar/tray (start / done / fail, real-time, styled after the main app HUD), macOS/Linux post a system notification on completion (NUPHUS_MCP_HUD=off to disable). Window activation is never used as a visibility fallback.

desktop_vision uses your own vision model. It speaks two protocols:

• OpenAI-compatible Chat Completions (default) — works with OpenAI, MiniMax, Qwen, Ollama, vLLM, …
• Anthropic native Messages API — point NUPHUS_MCP_VISION_BASE_URL at https://api.anthropic.com/v1 and the protocol is auto-detected from the host; or force it with NUPHUS_MCP_VISION_PROVIDER=anthropic.

Nothing is required unless you call this tool — and when it is not configured the tool returns a clear error instead of silently failing.

By default browser_* tools launch and manage their own Chrome instance. To drive an external browser instead — e.g. an anti-detect / fingerprint browser — start it with a debugging port and point the server at it:

# Example: start your fingerprint browser with a debugging port
chrome --remote-debugging-port=9222 --user-data-dir=...

// MCP client config
"env": { "NUPHUS_MCP_BROWSER_CDP_URL": "http://127.0.0.1:9222" }

When set, browser_* tools attach to that endpoint and never launch a managed Chrome; attach failures are hard errors (no silent fallback into the wrong browser). The external browser belongs to you — the server never kills it on exit.

Fingerprint browsers typically get a new random debug port every time a window is reopened, which would leave a fixed ..._CDP_URL pointing at a dead port. Provide the browser identity and the server re-resolves the live port automatically on the next tool call:

// MCP client config — attach + self-heal
"env": {
  "NUPHUS_MCP_BROWSER_CDP_URL": "http://127.0.0.1:9222",
  "NUPHUS_BROWSER_EXE_PATH": "C:\\path\\to\\fingerprint-browser.exe",
  "NUPHUS_BROWSER_NAME": "AdsPower",
  "NUPHUS_BROWSER_USER_DATA_DIR": "C:\\path\\to\\profile"
}

With the identity set, if the configured endpoint stops answering the server locates the window process by exe path, re-resolves its actual port (literal cmdline port, or DevToolsActivePort in the profile dir for random-port launches), verifies the candidate with a proxy-bypassing CDP probe and retries once — close and reopen the window and the next tool call just works. Without an identity, attach failures stay hard errors asking you to update the configured URL. Either way there is no fallback to a managed Chrome.

Want browser_* tools to drive a browser carrying your own extensions / bookmarks / login state? First, a hard Chrome 136+ restriction: after Chrome's official security change, --remote-debugging-port and --remote-debugging-pipe are ignored against the default user data directory — they must be paired with --user-data-dir pointing at a non-default directory. This is a deliberate design to stop infostealer malware from reading real cookies through the local debugging port — not a nuphus-mcp defect, and no flag / registry policy can bypass it (RemoteDebuggingAllowed only allows/disallows the switches; it cannot re-enable them on the default directory).

So "real default profile + CDP-controllable" are mutually exclusive since 136+. Pick one of the following, in priority order:

Option A (recommended, simplest): log in once in the nuphus-managed profile

nuphus-mcp manages its own Chrome instance by default (a dedicated --user-data-dir, always debuggable). Open it, log into the sites you want to stay signed in once — the state persists to that profile and every subsequent browser_* call carries those sessions. Zero config, zero copying.

Option B: copy your real profile to a debuggable directory (keeps extensions / bookmarks / logins)

When you need the original browser's extensions and login state, copy the real profile and launch from the copy:

1. Quit your running Chrome / Edge entirely first (profile-lock conflict)
2. Copy the profile to a non-default dir:
   • Windows: copy "%LOCALAPPDATA%\Google\Chrome\User Data\Default" <dest>\Default
   • macOS: cp -R ~/Library/"Application Support"/Google/Chrome/Default <dest>/Default
   • Linux: cp -R ~/.config/google-chrome/Default <dest>/Default
3. Launch the copy with a debug port: chrome --remote-debugging-port=9222 --user-data-dir=<dest>
4. Attach as an external browser: NUPHUS_MCP_BROWSER_CDP_URL=http://127.0.0.1:9222 (add NUPHUS_BROWSER_EXE_PATH / NUPHUS_BROWSER_USER_DATA_DIR for port self-healing)

Notes: on Windows, cookies / passwords are DPAPI user-level encrypted and decrypt fine from the copy under the same user — login state mostly survives; on macOS some credentials live in Keychain and a few sites may ask to log in again. The copy can be hundreds of MB to GBs, and it cannot run at the same time as the real browser.

Option C: desktop visual automation — operate the actually-running browser window

To operate the user's currently running real browser (DOM-level is not possible), use the desktop OCR + mouse/keyboard chain (desktop_perceive / desktop_* tools) — visually click, type and read the screen without CDP, at the cost of no DOM access and no login-state injection.

desktop_perceive runs PaddleOCR and YOLO icon detection locally with ONNX Runtime. The first call downloads the OCR models and icon_detect.onnx together automatically into %APPDATA%\Nuphus\models (or NUPHUS_MODELS_DIR). Download failures return a clear error with manual instructions. YOLO is optional at runtime: if its download fails, perceive still returns OCR elements and reports yolo_available: false (use NUPHUS_MCP_YOLO_MODEL_URL for a custom source). See TOOLS.md → Vision & Local Models.

All other tools need no API key.

Install via npm (recommended — all platforms, prebuilt binaries):

npm install -g @nuphus/nuphus-mcp

The nuphus-mcp meta package installs the prebuilt binary for your platform automatically (Windows x64/arm64, macOS arm64, Linux x64/arm64) and puts the nuphus-mcp command on your PATH. No Rust toolchain needed:

nuphus-mcp   # stdio MCP server

Build from source (requires the Rust toolchain):

cargo build --release -p nuphus-mcp
# binary at target/release/nuphus-mcp(.exe)

The server reads newline-delimited JSON from stdin and writes JSON-RPC responses to stdout. Logs go to stderr.

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

This server can physically control the machine it runs on. By default write tools run without confirmation; we strongly recommend enabling strict confirmation so destructive operations require an explicit "confirm": true argument from the client (otherwise the tool is rejected with isError).

Any one of the following:

# CLI flag
nuphus-mcp --confirm-write

# Environment variable (recommended — survives across clients, keeps config simple)
export NUPHUS_MCP_CONFIRM_WRITE=1      # macOS / Linux
setx NUPHUS_MCP_CONFIRM_WRITE 1        # Windows (persistent for new shells)

# MCP client args
"args": ["--confirm-write"]

Claude Desktop — recommended claude_desktop_config.json:

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

Prefer the environment variable: one setting applies to every MCP client on the machine. See SECURITY.md and the Safety Annotations section of TOOLS.md for the full threat model.

Point any MCP client at nuphus-mcp. After npm install -g @nuphus/nuphus-mcp the command is on your PATH; otherwise use the absolute path to the binary (nuphus-mcp / nuphus-mcp.exe).

Claude Desktop — claude_desktop_config.json:

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

Any MCP client (generic mcpServers JSON):

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

Supported MCP methods: initialize, notifications/initialized, ping, tools/list, tools/call.

Recommended — install the official dsh-nuphus-mcp plugin (Gitee mirror: gitee.com/nuphus/dsh-nuphus-mcp). It mounts nuphus-mcp into DSH as native tools with zero config — --confirm-write on by default, no code change needed:

npx -p @deepseek-ai/dsh dsh plugin --profile web add github:mrpulor-gh/dsh-nuphus-mcp

Manual alternative — plug into DeepSeek Harness as a plain stdio MCP server via its built-in MCP client (@deepseek-ai/dsh-mcp-client). Mount in DSH's cordis.yml / patch:

- id: nuphus-mcp
  name: '@deepseek-ai/dsh-mcp-client'
  config:
    serverName: nuphus-mcp
    transport: stdio
    command: nuphus-mcp
    args: ["--confirm-write"]
    toolCallTimeoutMs: 120000   # DSH default 60000 is too low for screenshots/OCR

Tools register as mcp__nuphus-mcp__* (e.g. mcp__nuphus-mcp__desktop_click). Run DSH in the desktop session of the machine you want controlled.

A self-contained stdio client that walks through initialize → tools/list → tools/call:

cargo build -p nuphus-mcp
cargo run -p nuphus-mcp --example demo

cargo check --workspace
cargo test -p nuphus-mcp          # protocol + security + vision + models tests (60)

This server can physically control the machine it runs on. Read SECURITY.md and the Safety Annotations section of TOOLS.md before deploying. Recommended: run with --confirm-write (or NUPHUS_MCP_CONFIRM_WRITE=1) so write tools require an explicit "confirm": true argument.

Thanks to the people who have contributed code to this project:

MIT