pinchtab-mcp
MCP server that wraps PinchTab's HTTP API for browser automation, enabling AI agents to control Chrome via actions like navigation, clicking, typing, and screenshots.
README
<h1 align="center">pinchtab-mcp</h1>
<p align="center"> <strong>MCP server for <a href="https://github.com/pinchtab/pinchtab">PinchTab</a> — browser automation for AI agents</strong><br/> Works with OpenCode, Cursor, Claude Desktop, and any MCP-compatible client </p>
<p align="center"> <img src="https://img.shields.io/badge/MCP-stdio-6B46C1?style=flat-square&logo=data:image/svg+xml;base64,PHN2ZyB4bWxucz0iaHR0cDovL3d3dy53My5vcmcvMjAwMC9zdmciIHZpZXdCb3g9IjAgMCAyNCAyNCI+PHBhdGggZmlsbD0id2hpdGUiIGQ9Ik0xMiAyQzYuNDggMiAyIDYuNDggMiAxMnM0LjQ4IDEwIDEwIDEwIDEwLTQuNDggMTAtMTBTMTcuNTIgMiAxMiAyem0tMSAxNXYtNEg3bDUtOXY0aDRsLTUgOXoiLz48L3N2Zz4=" alt="MCP stdio"/> <img src="https://img.shields.io/badge/node-%3E%3D18-339933?style=flat-square&logo=node.js&logoColor=white" alt="Node.js ≥18"/> <img src="https://img.shields.io/badge/TypeScript-5.x-3178C6?style=flat-square&logo=typescript&logoColor=white" alt="TypeScript"/> <img src="https://img.shields.io/badge/license-MIT-blue?style=flat-square" alt="MIT license"/> </p>
What is this?
PinchTab is a standalone Go binary that gives AI agents full control over a Chrome browser via an HTTP API. It is token-efficient, headless-capable, and supports persistent browser profiles.
pinchtab-mcp is a thin Model Context Protocol (MCP) stdio server that wraps PinchTab's HTTP API — making it available as a standard MCP tool in any compatible AI coding agent or chat client.
AI client (OpenCode / Cursor / Claude Desktop)
│ MCP stdio (JSON-RPC)
▼
pinchtab-mcp ──HTTP──▶ PinchTab :9867 ──CDP──▶ Chrome
Why a separate MCP wrapper?
PinchTab exposes a plain HTTP API. MCP clients communicate over stdin/stdout using JSON-RPC. This server bridges the two, adding:
- Single
pinchtabtool with a unifiedactionparameter — minimal context bloat - Typed, validated inputs via Zod
- Auth token forwarding, configurable timeout
- Screenshot responses as MCP image content (base64 JPEG)
Prerequisites
| Requirement | Notes |
|---|---|
| Node.js ≥ 18 | Required to run the MCP server |
| PinchTab | The Go binary must be running locally (or in Docker) |
Install PinchTab
# macOS / Linux — recommended
curl -fsSL https://pinchtab.com/install.sh | bash
# Docker
docker run -d -p 9867:9867 ghcr.io/pinchtab/pinchtab:latest
Note:
npm install -g pinchtabdoes not reliably install the binary on all platforms. Use the install script or Docker instead.
Full PinchTab docs: pinchtab.com/docs
Setup
1. Clone and build
git clone https://github.com/domci/pinchtab-mcp.git
cd pinchtab-mcp
npm install
npm run build
This compiles src/index.ts to dist/index.js.
2. Start PinchTab
In a separate terminal — PinchTab must be running before the MCP server is used:
# Basic
pinchtab
# With an auth token (recommended)
BRIDGE_TOKEN=my-secret pinchtab
If the MCP tool returns
Connection failed, PinchTab is not running. Start it as above and retry.
3. Configure your client
OpenCode
Add to ~/.config/opencode/opencode.json (global) or opencode.json in your project root:
{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"pinchtab": {
"type": "local",
"command": ["node", "/absolute/path/to/pinchtab-mcp/dist/index.js"],
"enabled": true,
"environment": {
"PINCHTAB_URL": "http://localhost:9867",
"PINCHTAB_TOKEN": "my-secret" // omit if no auth token
}
}
}
}
Cursor IDE
Add to ~/.cursor/mcp.json (global) or .cursor/mcp.json in your project root:
{
"mcpServers": {
"pinchtab": {
"command": "node",
"args": ["/absolute/path/to/pinchtab-mcp/dist/index.js"],
"env": {
"PINCHTAB_URL": "http://localhost:9867",
"PINCHTAB_TOKEN": "my-secret"
},
"type": "stdio"
}
}
}
Claude Desktop
Add to ~/Library/Application Support/Claude/claude_desktop_config.json:
{
"mcpServers": {
"pinchtab": {
"command": "node",
"args": ["/absolute/path/to/pinchtab-mcp/dist/index.js"],
"env": {
"PINCHTAB_URL": "http://localhost:9867"
}
}
}
}
Environment variables
| Variable | Default | Description |
|---|---|---|
PINCHTAB_URL |
http://localhost:9867 |
Base URL of the running PinchTab server |
PINCHTAB_TOKEN |
(empty) | Bearer token — must match BRIDGE_TOKEN set on PinchTab |
PINCHTAB_TIMEOUT |
30000 |
HTTP request timeout in milliseconds |
Tool reference
A single pinchtab tool is registered. All operations are dispatched via the action parameter.
Actions
| Action | Description | Key parameters |
|---|---|---|
navigate |
Navigate to a URL | url, newTab?, blockImages?, timeout? |
snapshot |
Accessibility tree of the current page | filter?, format?, diff?, maxTokens?, depth? |
click |
Click an element | ref |
type |
Type text into a focused element | ref, text |
fill |
Clear and fill an input | ref, text |
press |
Press a key | ref, key (e.g. Enter, Tab) — for form submission prefer click on the submit button |
hover |
Hover over an element | ref |
scroll |
Scroll the page | ref?, scrollY |
select |
Select a dropdown option | ref, value |
focus |
Focus an element | ref |
text |
Extract readable page text (~800 tokens) | mode? (readability|raw) |
tabs |
List, open, or close tabs | tabAction? (list|new|close) |
screenshot |
Capture a JPEG screenshot | quality? (1–100) |
evaluate |
Execute JavaScript in the page | expression |
pdf |
Export page as PDF | landscape?, scale? |
health |
Check PinchTab connectivity | — |
All actions accept an optional tabId to target a specific tab.
Token strategy
| Scenario | Recommended action | Approx. tokens |
|---|---|---|
| Read page content | text |
~800 |
| Find interactive elements | snapshot with filter=interactive&format=compact |
~3,600 |
| Track page changes | snapshot with diff=true |
delta only |
| Visual verification | screenshot |
~2,000 |
Example prompts
Navigate to https://news.ycombinator.com and extract the top 10 story titles.
use pinchtab
Go to https://example.com/login, fill in the username and password fields, and submit the form.
use pinchtab
Take a screenshot of the current page.
use pinchtab
Development
# Install dependencies
npm install
# Build (TypeScript → dist/)
npm run build
# Run directly
node dist/index.js
Project structure
pinchtab-mcp/
├── src/
│ └── index.ts # MCP stdio server — all logic lives here
├── dist/ # Compiled output (git-ignored)
├── package.json
└── tsconfig.json
Troubleshooting
Connection failed: fetch failed. Is PinchTab running at http://localhost:9867?
The PinchTab binary is not running. Start it with pinchtab in a separate terminal, then retry.
npm install -g pinchtab installs but pinchtab command is not found
The npm package does not install the Go binary on all platforms. Use the install script instead:
curl -fsSL https://pinchtab.com/install.sh | bash
press Enter doesn't submit a form / page doesn't navigate
Some sites handle form submission via click events on the submit button rather than keyboard events on the input. Use click on the submit button ref instead of press Enter on the input.
Search input shows "queryEnter" as its value
This happens when press appends literally to the field value. Use fill to set the value cleanly, then click the submit button.
Security notes
BRIDGE_TOKEN/PINCHTAB_TOKEN— always set a token in production environments; rotate it regularly.evaluateexecutes arbitrary JavaScript inside Chrome — restrict access to trusted agents and domains.- PinchTab should not be exposed to the public internet; keep it on
localhostor behind a private network.
License
MIT — see LICENSE.
<p align="center"> Built on top of <a href="https://github.com/pinchtab/pinchtab"><strong>PinchTab</strong></a> by the PinchTab authors · MCP wrapper by <a href="https://github.com/domci">domci</a> </p>
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。