mario-playwright-mcp
MCP server for browser automation with AI agents, enabling network payload capture and QA workflows.
README
Playwright MCP Mario
Browser MCP for AI agents — based on Microsoft Playwright MCP with extensions for network payload capture and QA workflows.
What is this?
An MCP server that lets LLMs control a browser through accessibility snapshots (no vision models needed). This fork adds network payload capture and sensitive redaction — better for API debugging and QA.
flowchart LR
subgraph Card1["1- Default Playwright"]
D1[Microsoft upstream]
end
subgraph Card2["2- Mario Playwright"]
D2[network payloads · redaction]
end
subgraph Card3["3- Mario agent skills"]
D3[qa-agent · QA workflows]
end
subgraph Runtime[runtime]
direction LR
subgraph BrowserCard["4- Browser"]
Browser[Chromium]
end
subgraph ToolsCard["5- Tools"]
T1[navigate]
T2[click · type]
T3[snapshot · screenshot]
T4[network ++]
end
end
Card1 --> Card2 --> Card3 --> Runtime
BrowserCard --> ToolsCard
- MCP = bridge that controls the browser and returns snapshots, screenshots, console, network.
- Recommended: qa-agent skill for QA workflows.
- Mario vs upstream: full payloads, URL filter, sensitive redaction — not just
[POST] url => [200].
Quick start
1. Clone and install
git clone https://github.com/YOUR_ORG/mario-playwright-mcp.git
cd mario-playwright-mcp
npm install
npx playwright install chromium
2. Configure your MCP client
Cursor (in ~/.cursor/mcp.json or Settings → MCP):
{
"mcpServers": {
"playwright-mcp": {
"command": "node",
"args": [
"/path/to/mario-playwright-mcp/packages/playwright-mcp/cli.js",
"--output-dir",
".mcp-output"
]
}
}
}
Output (console logs, screenshots, etc.) goes to .mcp-output/ in the workspace. Add .mcp-output/ to .gitignore to keep the repo clean.
Generic config (VS Code, Claude Desktop, etc.) — use node + path to cli.js as above.
3. Use it
Open a project, start a chat, and use tools like browser_navigate, browser_snapshot, browser_network_requests.
QA Agent skill (recommended)
For QA workflows and structured test reports, use the qa-agent skill — an agent skill for QA engineers that uses this MCP for browser testing, post-action analysis (console + network), and evidence capture.
Tools
All tools the Playwright MCP exposes. ✨ = Mario-enhanced (better in this fork).
Core (always available)
| Tool | Description |
|---|---|
browser_navigate |
Navigate to a URL |
browser_navigate_back |
Go back |
browser_snapshot |
Capture accessibility tree (better than screenshot for actions) |
browser_take_screenshot |
Take a screenshot |
browser_console_messages |
Get console logs |
browser_network_requests ✨ |
List network requests — Mario: adds includePayloads, url filter, maxBodySize, sensitive redaction |
browser_click |
Click an element |
browser_type |
Type text into an element |
browser_hover |
Hover over an element |
browser_drag |
Drag and drop |
browser_fill_form |
Fill multiple form fields |
browser_select_option |
Select dropdown option |
browser_press_key |
Press a key |
browser_resize |
Set viewport size — browser opens at native resolution by default; use to test Mobile (375×667), Tablet (768×1024), Desktop (1920×1080) |
browser_evaluate |
Run JavaScript on the page |
browser_run_code |
Run Playwright code snippet |
browser_file_upload |
Upload files |
browser_handle_dialog |
Accept/dismiss dialogs |
browser_wait_for |
Wait for text or time |
browser_tabs |
List, create, close, or switch tabs |
browser_close |
Close the browser |
browser_install |
Install the browser (if missing) |
Opt-in (pass --caps=…)
| Capability | Tools |
|---|---|
--caps=pdf |
browser_pdf_save — Save page as PDF |
--caps=vision |
browser_mouse_move_xy, browser_mouse_click_xy, browser_mouse_drag_xy, browser_mouse_down, browser_mouse_up, browser_mouse_wheel — Coordinate-based actions |
--caps=testing |
browser_generate_locator, browser_verify_element_visible, browser_verify_list_visible, browser_verify_text_visible, browser_verify_value — Test assertions |
Extensions (vs upstream)
Network payload capture
browser_network_requests supports:
| Param | Description |
|---|---|
includePayloads |
When true, include request/response headers and bodies |
url |
Filter by URL substring (e.g. /login, /api/users) |
maxBodySize |
Max response body size (default 50KB) |
includeStatic |
Include images, fonts, etc. (default false) |
Example — get full details for a login call:
browser_network_requests(url: "/login", includePayloads: true)
Output:
[POST] https://api.example.com/login => [200] OK
Request headers: { "content-type": "application/json", ... }
Request body: {"email":"user@example.com","password":"***"}
Response headers: { "content-type": "application/json", ... }
Response body: {"token":"***","user":{...}}
Sensitive keys (password, token, secret, authorization, cookie, api_key) are redacted as ***.
Workflow
browser_network_requests(includeStatic: false)→ list URLsbrowser_network_requests(url: "/api/endpoint", includePayloads: true)→ headers + bodies for that request
Project structure
mario-playwright-mcp/
├── packages/
│ ├── playwright-mcp/ # MCP server
│ └── extension/ # Browser extension (upstream)
├── patches/ # Network payload patch for Playwright
├── docs/ # Extra docs
│ └── UPSTREAM.md # Update from Microsoft upstream
└── README.md
Update from upstream
git remote add upstream https://github.com/microsoft/playwright-mcp.git
git fetch upstream
git merge upstream/main
npm install
If Playwright version changes and the patch fails, re-apply changes in node_modules/playwright/lib/mcp/browser/tools/network.js and run npx patch-package playwright. See docs/UPSTREAM.md.
Requirements
- Node.js 18+
- MCP client (Cursor, VS Code, Claude Desktop, Goose, etc.)
For client-specific configs (Amp, Cline, LM Studio, etc.) and CLI options, see the upstream README.
Versioning
- main — development branch
- 1.1 — first Mario fork release (network payload capture, .mcp-output, etc.)
- 1.2, 1.3, … — future releases as changes land
Use semantic versioning: patch for fixes (1.1.1), minor for features (1.2.0).
License
Apache 2.0 — see LICENSE. Based on Microsoft Playwright 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。