cloakbrowser-mcp-server
Enables human-in-the-loop monitoring of browser automation by providing a VNC view of the Chromium browser driven by MCP, while exposing the MCP server over Streamable HTTP for client integration.
README
cloakbrowser-mcp-server
Headed wrapper around swimmwatch/cloakbrowser-mcp that exposes the Chromium browser over VNC so a human can watch the same session an MCP client is driving over Streamable HTTP.
The whole project is one Python launcher + one Dockerfile. It starts Xvfb + openbox so Chromium runs headed (required by anti-bot probes that inspect window/widget state), mirrors the display with x11vnc, then execs the upstream cloakbrowser-mcp CLI in Streamable HTTP mode.
Quick start
docker build -t cloakbrowser-mcp-server .
docker run --rm \
-p 3000:3000 \
-p 5900:5900 \
-e VNC_PASSWORD=changeme \
-e DISPLAY_WIDTH=1024 \
-e DISPLAY_HEIGHT=768 \
cloakbrowser-mcp-server
Then:
- MCP client — point it at
http://localhost:3000/mcp(any Streamable HTTP MCP client: Claude Desktop viaclaude mcp add --transport http cloakbrowser http://localhost:3000/mcp, Codex CLI, Cursor, etc.) - VNC viewer — connect to
localhost:5900, passwordchangeme. macOS built-in:open vnc://localhost:5900. Any VNC client works.
Architecture
┌────────────────────────────────────────────┐
│ container │
host ── 3000 ──► │ cloakbrowser-mcp (Streamable HTTP) │
│ └─ Playwright MCP ──► Chromium headed │
host ── 5900 ──► │ Xvfb :99 + openbox ◄── x11vnc mirror │
│ │
│ /data (bind-mounted, persistent profile) │
└────────────────────────────────────────────┘
scripts/launcher.py orchestrates the whole thing in one process:
Xvfb :99at$DISPLAY_WIDTH x $DISPLAY_HEIGHTopenbox(window manager so Chromium honours--start-maximized)x11vncon:5900, password from$VNC_PASSWORDor openexecnode /opt/cloakbrowser-mcp/dist/cli.js --transport streamable-http --http-host 0.0.0.0 --http-port 3000- force
PLAYWRIGHT_MCP_HEADLESS=falsein the child env (the wrapper's whole reason to exist) - wait for SIGTERM/SIGINT, tear down X + VNC
Configuration
Environment variables (wrapper-specific)
| Var | Default | Effect |
|---|---|---|
VNC_PASSWORD |
unset | If set, VNC requires this plaintext password. If unset, VNC is unauthenticated (x11vnc -nopw). |
DISPLAY_WIDTH |
1024 |
Xvfb screen width. Also passed to Playwright as the page viewport (unless the operator sets CLOAK_PLAYWRIGHT_MCP_CONTEXT_OPTIONS). |
DISPLAY_HEIGHT |
768 |
Xvfb screen height. Also passed to Playwright as the page viewport. |
Pass-through
Any PLAYWRIGHT_MCP_* and CLOAK_PLAYWRIGHT_MCP_* variable is forwarded to the upstream cloakbrowser-mcp CLI untouched. See Configuration in the upstream docs for the full list. The wrapper only overrides:
PLAYWRIGHT_MCP_HEADLESS=false(always)--transport streamable-http --http-host 0.0.0.0 --http-port 3000(always)
MCP auth, persistent profiles, Chrome extensions, regional proxies, humanized input — all configured via upstream env vars.
Hardcoded (cannot be changed via env)
| Thing | Value | Reason |
|---|---|---|
| MCP transport | streamable-http |
The wrapper exists to expose the MCP server over HTTP, not stdio. |
| MCP HTTP host | 0.0.0.0 |
Containerised; operator maps ports on docker run. |
| MCP HTTP port | 3000 |
Same. |
| MCP endpoint | /mcp |
Upstream default. |
| Display | :99 |
Xvfb + WM + Chromium all agree. |
| VNC port | 5900 |
Operator maps with -p. |
Volume
Bind-mount /data to persist Chromium's profile across container runs:
docker run --rm \
-p 3000:3000 -p 5900:5900 \
-v ~/cloak-mcp-data:/data \
-e VNC_PASSWORD=changeme \
cloakbrowser-mcp-server
Without the mount, /data lives inside the container and is lost on docker rm.
Port mapping
The internal MCP port is always 3000 and the internal VNC port is always 5900, regardless of what you map them to on the host:
# default: same ports on host and container
-p 3000:3000 -p 5900:5900
# remap both
-p 8080:3000 -p 5999:5900
Project layout
cloakbrowser-mcp-server/
├── Dockerfile # FROM swimmwatch/cloakbrowser-mcp:latest + X stack
├── README.md
├── pyproject.toml # ruff config (lint + format)
├── scripts/
│ └── launcher.py # single-file orchestrator
└── tests/
└── test_launcher.py # stdlib unittest, no test framework dependency
No build system, no CI — keep it boring.
Lint & test
ruff check . # lint
ruff format --check . # format check
ruff format . # format fix
python3 -m unittest discover -s tests -v # 12 unit tests, stdlib only
docker build -t cloakbrowser-mcp-server . # full image build (~30s with cached base)
License
Inherits MIT from cloakbrowser-mcp.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。