MIT 协议的跨平台桌面自动化 MCP 服务器：通过官方 dsh-nuphus-mcp 插件把 38 个 stdio 工具以原生方式挂载进 DeepSeek Harness，内置本地 OCR 与 BYOK 视觉能力。
DSH 适配
原生运行时
作者声明
安全审计
未审计
最后核验
2026-08-21
许可证
MIT
01它能帮你完成什么？
Give DeepSeek Harness computer-use MCP tools to control the desktop and Chrome
38 MCP tools (15 desktop + 23 browser) registered in DSH as `mcp__nuphus-mcp__*` — screenshots, window control, mouse/keyboard, Chrome CDP automation
Developers running DeepSeek Harness (dsh) who want their agent to operate the screen, windows and browser hands-free
Automate Chrome workflows from DSH sessions
Browser automation: navigate, snapshot (accessibility tree with `@N` refs), click, type, scroll, extract, screenshot, cookies get/set/import, tabs and downloads
Automation engineers who want an AI agent inside DeepSeek Harness to drive their own Chrome or an external fingerprint browser via CDP
Ground agent actions in what's on screen with vision and local OCR
`desktop_vision` (BYOK screenshot analysis via OpenAI-compatible or Anthropic native API) plus `desktop_perceive` (local PaddleOCR, optional YOLO icon detection) — semantic understanding plus pixel-precise coordinates
Teams building computer-use agents in DSH that need reliable element coordinates without sending screenshots to a hosted vision service
02如何接入 DeepSeek Harness？
前置条件
- Chrome or Edge installed — required for browser tools (the server auto-detects an installed browser)
- macOS only: Accessibility permission (System Settings → Privacy & Security → Accessibility) for desktop input
- Rust toolchain (stable) only if building from source with Cargo — not needed for the npm install path
安装步骤
- 01
Install via npm (recommended, all platforms, prebuilt binaries): `npm install -g @nuphus/nuphus-mcp` — puts the `nuphus-mcp` command on your PATH, no Rust toolchain needed
$ npm install -g @nuphus/nuphus-mcp
- 02
Recommended for DSH — install the official plugin: `npx -p @deepseek-ai/dsh dsh plugin --profile web add github:mrpulor-gh/dsh-nuphus-mcp`
$ npx -p @deepseek-ai/dsh dsh plugin --profile web add github:mrpulor-gh/dsh-nuphus-mcp
- 03
Manual alternative — mount in DSH's `cordis.yml` as a stdio MCP server via `@deepseek-ai/dsh-mcp-client` (command: nuphus-mcp, args: ["--confirm-write"], toolCallTimeoutMs: 120000)
- 04
Recommended: enable strict confirmation with `--confirm-write` or `NUPHUS_MCP_CONFIRM_WRITE=1` so destructive tools require an explicit `"confirm": true` argument
验证接入成功
- Quick smoke test: `echo '{"jsonrpc":"2.0","id":0,"method":"initialize",...}' | nuphus-mcp` — the server reads newline-delimited JSON from stdin and writes JSON-RPC responses to stdout
03DSH 适配与能力边界
Official dsh-nuphus-mcp plugin mounts nuphus-mcp into DeepSeek Harness as native tools with zero config (--confirm-write on by default); manual alternative via DSH's built-in stdio MCP client
38 MCP tools over stdio (15 desktop + 23 browser)
JSON-RPC 2.0 requests from any MCP client over stdin→screenshots, window control, mouse/keyboard, clipboard, Chrome CDP automation — one binary, no daemon, no network service
tools physically control screen, windows, mouse and keyboard on the host machineComputer vision pair (desktop_vision + desktop_perceive)
screenshot of the current screen→semantic analysis from your own vision LLM (BYOK) plus local PaddleOCR text elements with pixel coordinates; optional YOLO icon detection
desktop_vision sends screenshots to your own vision LLM endpoint (BYOK API key required)first perceive call downloads OCR/YOLO models into %APPDATA%\Nuphus\models (or NUPHUS_MODELS_DIR)Chrome automation via CDP (managed or external browser)
browser tool calls; optionally NUPHUS_MCP_BROWSER_CDP_URL to attach to an external/fingerprint browser→navigate, click, type, extract, screenshot, cookies, tabs, downloads — with self-healing port re-resolution for fingerprint browsers
launches and manages a Chrome instance by default; attaches to your external browser when NUPHUS_MCP_BROWSER_CDP_URL is set (never kills it on exit)Safety-first write controls
--confirm-write flag or NUPHUS_MCP_CONFIRM_WRITE=1→destructive tools require an explicit "confirm": true argument, otherwise the tool is rejected with isError; path validation for screenshots, uploads and file drags
04适合谁？何时不该用？
适合
- Developers running DeepSeek Harness (dsh) who want their agent to operate the screen, windows and browser hands-free
- Automation engineers who want an AI agent inside DeepSeek Harness to drive their own Chrome or an external fingerprint browser via CDP
- Teams building computer-use agents in DSH that need reliable element coordinates without sending screenshots to a hosted vision service
不适合
- Desktop control is full on Windows (Win32); on macOS desktop input requires Accessibility permission; on Linux window/input capabilities are partial.
- The server can physically control the machine it runs on; by default write tools run without confirmation, so enabling --confirm-write (or NUPHUS_MCP_CONFIRM_WRITE=1) is strongly recommended.
05兼容性、维护与安全提示
- Desktop control is full on Windows (Win32); on macOS desktop input requires Accessibility permission; on Linux window/input capabilities are partial.
- The server can physically control the machine it runs on; by default write tools run without confirmation, so enabling --confirm-write (or NUPHUS_MCP_CONFIRM_WRITE=1) is strongly recommended.
- Desktop & browser automation need no API key, but desktop_vision requires a BYOK vision endpoint, and the first desktop_perceive call downloads OCR/YOLO models from the network.
MIT · actively maintained (latest release v0.1.12, 2026-08-17)
06常见问题
如何把 nuphus-mcp 接入 DeepSeek Harness？
推荐安装官方 dsh-nuphus-mcp 插件：`npx -p @deepseek-ai/dsh dsh plugin --profile web add github:mrpulor-gh/dsh-nuphus-mcp`。它会把 nuphus-mcp 以原生工具形式零配置挂载进 DSH，且默认开启 --confirm-write。也可以在 DSH 的 cordis.yml 中通过内置的 `@deepseek-ai/dsh-mcp-client` 手动挂载为 stdio MCP 服务器。
DSH 用的是原生工具还是普通 MCP 模式？
两种都可以。官方插件把 38 个工具原生挂载进 DSH；手动方式则把 nuphus-mcp 作为普通 stdio MCP 服务器接入，工具注册为 `mcp__nuphus-mcp__*`。无论哪种方式，都请在你想控制的机器上的桌面会话中运行 DSH。
需要 API key 吗？
桌面和浏览器自动化都不需要。只有 `desktop_vision` 需要你自带视觉模型（BYOK，通过 NUPHUS_MCP_VISION_* 环境变量支持 OpenAI 兼容或 Anthropic 原生协议）；`desktop_perceive` 在本地运行 PaddleOCR，首次调用时自动下载模型。
支持哪些平台？
浏览器工具在 Windows 和 macOS 上功能完整，Linux 可用。桌面工具在 Windows（Win32）上功能完整，macOS 需授予辅助功能权限，Linux 为部分支持。npm 提供 Windows x64/arm64、macOS arm64 和 Linux x64/arm64 的预编译二进制。
如何保证写操作的安全？
该服务器可以物理控制它所在的机器。使用 `--confirm-write` 或 `NUPHUS_MCP_CONFIRM_WRITE=1` 开启严格确认后，破坏性工具必须携带显式的 "confirm": true 参数才会执行；dsh 插件默认已开启 --confirm-write。
07相关的 DSH 工作流
ruflo
作者 ruvnet
原版智能体元框架（meta-harness）：支持多智能体集群部署、自主工作流编排与对话式 AI 系统构建，内置自适应记忆、自学习智能与 RAG 检索，原生集成 Claude Code / Codex / Hermes 等大量工具。
reactive-resume
作者 amruthpillai
注重隐私的开源简历生成器，安全、可定制、可移植，永久免费，支持自托管。
everos
作者 evermind-ai
为所有 AI 智能体提供的便携统一记忆层，本地优先、Markdown 原生、数据归用户所有，跨应用与工作流自进化。
yao
作者 yaoapp
在桌面、移动端、浏览器或 API 上统一管理你的智能体与工作区，看板式任务追踪，支持自托管。
08数据与来源
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
页面基于项目公开文档、仓库元数据和 DSH Plugins 的结构化解析生成；最后核验于 2026-08-21。发现错误？提交更正。
最佳 DeepSeek Harness 插件
从全目录挑出的 12 个值得优先安装的插件，覆盖各个分类。