Free guide

Lightpanda: The Headless Browser Built for AI Agents, Explained

An open-source browser written from scratch in Zig that reads pages for agents and never draws them. What it is, how it works, how to set it up, and the catch.

By Beau Johnson·
The Short this guide goes with: Lightpanda: The Headless Browser Built for AI Agents, Explained

Quick answer

Lightpanda is an open-source (AGPL-3.0) headless browser built from scratch for AI agents and automation. It fetches pages, builds the DOM and runs their JavaScript in V8, but it has no layout or paint engine, so its screenshots are a picture of the page's text. On a Mac, install it with brew install lightpanda-io/browser/lightpanda, then point your Puppeteer or Playwright script at its CDP server or add it to Claude Code as an MCP server.

The short version

  • Lightpanda is an open-source headless browser written in Zig. It is not a Chromium fork, and it had over 41,000 GitHub stars on November 10, 2026.
  • It skips layout and paint. In Lightpanda's own test (25 at once, 933 pages of its own demo site) it used 123 MB of memory against Chrome's 2 GB and finished in 4.81 s against 46.70 s.
  • It is for agents and scripts that read the page and click things. If your agent works from screenshots, it won't help.

How it actually works

A browser is a stack

A normal browser does two jobs. The bottom of the stack fetches the page, parses the HTML into a DOM and runs the JavaScript. The top lays everything out and paints every pixel for your eyes. Lightpanda keeps the bottom (libcurl for fetching, html5ever for parsing, V8 for JavaScript) and has no graphical rendering engine at all.

A telegram, not a magazine

A magazine is the words plus the layout and the photos. A telegram is just the words, and it gets there faster. An agent that reads the page and clicks buttons only needs the telegram, so Lightpanda never spends memory or time on the magazine.

Why its screenshot is a picture of words

Since version 0.4.0 it can answer a screenshot request, but its source code says what that is: a text-only rasterizer. With no layout engine, a screenshot is the page's text flowed into blocks, the same content the markdown dump gives you. We ran it on example.com: no layout, no images, just the words.

It talks to the tools you already use

It runs a CDP server, so an existing Puppeteer script connects by changing its browserWSEndpoint, and Playwright can connect over CDP too. WebDriver BiDi is supported, and lightpanda mcp starts a native MCP server for agents like Claude Code.

How to set it up

  1. 1.Install it on a Mac

    Homebrew installs the latest nightly. On Linux, download the binary from the releases page. There is no native Windows build: install the Linux version inside WSL.

    brew install lightpanda-io/browser/lightpanda
  2. 2.Turn telemetry off (optional)

    Lightpanda collects and sends usage telemetry by default. This environment variable turns it off.

    export LIGHTPANDA_DISABLE_TELEMETRY=true
  3. 3.Read a page as markdown

    The quickest first run. It loads the page, runs its JavaScript and prints the content.

    lightpanda fetch --obey-robots --dump markdown https://example.com
  4. 4.Point Puppeteer or Playwright at it

    Start the CDP server, then connect your script to ws://127.0.0.1:9222 instead of launching Chrome.

    lightpanda serve --obey-robots --host 127.0.0.1 --port 9222
    # Puppeteer: puppeteer.connect({ browserWSEndpoint: "ws://127.0.0.1:9222" })
  5. 5.Or give it to Claude Code as an MCP server

    Add it to your MCP configuration with the full path to the binary, so your agent can browse through Lightpanda.

    { "mcpServers": { "lightpanda": { "command": "/path/to/lightpanda", "args": ["mcp"] } } }

The fine print

  • License: AGPL-3.0. You can use it for free, including at work. If you modify Lightpanda and let people use your modified version over a network, the AGPL says you must offer them your source code. Contributions need a CLA.
  • The speed and memory numbers are Lightpanda's own benchmark on its own demo site, 25 processes against 25 Chrome tabs. One process against one tab, its numbers show about 1.6x faster. Test it on your own pages.
  • No layout means anything that depends on how the page looks can break: visual click targets, canvas apps, and agents that read screenshots.
  • Telemetry is on by default. Set LIGHTPANDA_DISABLE_TELEMETRY=true to turn it off.
  • No native Windows build (use WSL). It is not Chrome, so not every site works the same way.

FAQ

Is Lightpanda free?

Yes. It is open source under the AGPL-3.0 license and free to download and use. If you modify it and offer your modified version to others over a network, you have to share your changes under the same license.

Can Lightpanda take screenshots?

Sort of. Since version 0.4.0 it answers screenshot requests, but it has no layout engine, so the image is the page's text flowed into blocks, not what the page looks like in Chrome.

Is Lightpanda really 9x faster than Chrome?

That is Lightpanda's own benchmark: 25 Lightpanda processes against 25 Chrome tabs crawling 933 pages of its own demo site took 4.81 s against 46.70 s, using 123 MB against 2 GB. One against one, its numbers show about 1.6x faster. Your pages may differ.

Does Lightpanda work with Puppeteer and Playwright?

Yes, over CDP. Start lightpanda serve and connect your script to ws://127.0.0.1:9222 instead of launching Chrome.

Does Lightpanda run on Windows?

Not natively. The README says to install the Linux build inside WSL. Your Puppeteer or Playwright script can still run on the Windows side.

Sources

Facts checked 2026-11-10 against the Lightpanda README, LICENSE, src/rust/render/lib.rs, the demo repo's BENCHMARKS.md, the GitHub API (41,742 stars) and our own sandbox run of version 1.0.3 on November 9.

Want help building with tools like this?

Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.

Join Shipping Skool

More free guides