PinchTab CLI
Chrome control for AI agents: a local Go HTTP server with a CLI, MCP, profiles and a skill for Claude Code
- 10,577 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The agent gets control of a real Chrome together with profiles, cookies and saved logins
- Fingerprint masking and getting past site checks may breach those sites' terms and the law
- A port 9867 or MCP exposed outward gives an outsider full control of the browser
Install
- Works in
Manual install
curl -fsSL https://pinchtab.com/install.sh | bashThe install script from the README. Read it before running; Homebrew or npm work without a script.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The agent gets control of a real Chrome together with profiles, cookies and saved logins
- Fingerprint masking and getting past site checks may breach those sites' terms and the law
- A port 9867 or MCP exposed outward gives an outsider full control of the browser
This is third-party code. Review the repository files before installing.
What it does
PinchTab is a Go HTTP server of about 30 MB that controls an installed Chrome or Chromium and hands the agent a compact description of the page. Page text takes around 800 tokens, 5 to 13 times cheaper than screenshots, and elements get stable refs such as `e5` that live until navigation to a new document. The terminal commands `pinchtab nav`, `snap`, `click` and `text` drive the browser, and the same is available through an HTTP API with a token and through the `pinchtab mcp` MCP server. The server keeps several isolated Chrome instances with profiles: a profile stores logins, cookies and extensions, and the agent picks one by name. `pinchtab audit` crawls a site and builds a report with screenshots, console errors, broken assets, an accessibility score and Core Web Vitals, while `pinchtab compare` compares two versions of a site. The npm package copies the `pinchtab` skill into agent directories on its own.
Who it is for. Developers and testers who let Claude Code, Cursor or Codex drive a browser on their own machine and watch token spend.
Good fit when
- An agent browses documentation and sites, screenshots eat the token budget, and a text snapshot is cheaper
- A task needs a signed-in profile: the agent opens a work profile and downloads a weekly report
- Before a release the live site should be compared with staging and CI should stop on a difference
- A site audit from the sitemap with a `report.json` and screenshots is required
- Several agents run in parallel and each needs its own isolated Chrome instance
Not a fit when
- The server is meant to be exposed to the internet or to several users: it is designed for one operator on their own machine
- A ready cloud browser is wanted: you install and maintain PinchTab yourself
- A browser other than Chrome or Chromium is needed: PinchTab drives only those
Example request
Open our documentation page through PinchTab, take a structure snapshot and find where token sign-in is describedLimitations
The dashboard, HTTP API, MCP server and remote integrations are privileged control channels for a trusted operator and must not be exposed outward. By default the server listens on `127.0.0.1`, sensitive endpoint groups are off and the prompt-injection defense allows only locally hosted sites; widening that list knowingly lowers security. macOS and Linux are the best tested, Windows support is limited and `pinchtab server` should be run directly there. On macOS keep a separate Chrome for Testing for automation.
How to disable. Stop the background PinchTab process with your system tools and remove the install: `npm uninstall -g pinchtab` or `brew uninstall pinchtab`. Delete the skill from `~/.claude/skills`, `~/.cursor/skills` and other agent directories by hand.
FAQ
How do I install PinchTab?
Run `curl -fsSL https://pinchtab.com/install.sh | bash`, `brew install pinchtab/tap/pinchtab` or `npm install -g pinchtab`. Then `pinchtab daemon install` starts a background server on port 9867 with headless Chrome.
How do I use PinchTab as an MCP server?
Add a server to the client config with the command `pinchtab` and the argument `mcp`; for Claude Desktop, Cursor and VS Code that is three similar JSON blocks. `pinchtab server` or `pinchtab bridge` must be running first.
Why does the agent refuse to open external sites?
By default the prompt-injection defense lets it visit only locally hosted sites. You can widen the list in settings, but then hostile pages gain more room to influence the agent.
How does PinchTab save tokens?
It returns page text and the accessibility tree, around 800 tokens per page, instead of screenshots that cost 5 to 13 times more. Element refs such as `e5` survive a change of filter and depth.
Can it audit and compare sites?
Yes. `pinchtab audit` writes `report.json` and a screenshots folder, and `pinchtab compare` with the `--fail-on-diff` flag works as a CI check.
Discussion
No questions or reviews yet.
Related
CLI toolsAn agent drives your real logged-in Chrome: a CLI, MCP and an extension with a separate tab group per session
A skill for Amp, Claude Code and Cursor: the agent drives your live Chrome with its open tabs and signed-in accounts through a single debugging toggle
An MCP server and Chrome extension give any agent 22 tools for your own logged-in browser, verifying the effect of every action.
A CLI and skill for driving Chrome from an agent: page snapshots, clicks, console, network and audits in compact output
A CLI and MCP server that lets an agent drive your real Chrome with its logged-in sessions: 103 commands across 36 platforms plus browser automation
Console browser mb for agents: open a page, snapshot it, click, fill forms, take screenshots and record the screen
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| PinchTab (this card) | yes | 10,577 | ||
| agent-browser | yes | not required | 50,096 | a browser control CLI without site audits or version comparison |
| Playwright MCP | yes | not required | 39,570 | the Playwright MCP server without profiles, a daemon or token-saving snapshots |
README in short
The README presents PinchTab as a standalone server and not an extension, and stresses its local-first security model. It then walks through scenarios: headless and headed navigation, site audits, Docker isolation and distributed runs, followed by installation, the skill, and use through the CLI and HTTP API.