Lightpanda is a headless web browser written from scratch in Zig for AI agents, crawling, and web automation. It is not based on Chromium, Blink, or WebKit and has no graphical rendering engine. It executes JavaScript with the V8 engine, parses HTML with html5ever, and loads pages over libcurl, so it handles Ajax, single-page apps, and JavaScript frameworks that plain HTTP requests cannot.
| Metric (933-page crawl benchmark) | Lightpanda | Headless Chrome | Difference |
|---|---|---|---|
| Memory (peak, 100 pages) | 123MB | 2GB | ~16x less |
| Execution time (100 pages) | 5s | 46s | ~9x faster |
Existing automation clients connect to it as they would to Chrome: it serves the Chrome DevTools Protocol (CDP) and WebDriver BiDi, so Puppeteer and Playwright scripts run against it. A fetch command dumps a single URL as HTML, Markdown, PNG, or PDF, with options to wait for a selector or script before the dump. Lightpanda runs on Linux and macOS, inside WSL on Windows, and as a Docker image.
An agent mode drives the browser from plain-English instructions or slash commands inside the browser process. A session exports as a PandaScript, a plain JavaScript file with native browser primitives that replays deterministically without a model. Lightpanda also runs as an MCP server over stdio or HTTP, where each connection gets an isolated browsing session with its own page, cookies, and memory, and an agent skill is published separately. Usage telemetry is on by default and can be disabled.
Features
- Built from scratch: a Zig browser engine with V8 JavaScript, DOM APIs, XHR and Fetch, and no rendering overhead
- CDP and WebDriver BiDi: Puppeteer, Playwright, and other CDP clients connect without code changes beyond the endpoint
- Page dump: fetches a URL and outputs HTML, Markdown, PNG, or PDF
- Agent mode: a native agent navigates, clicks, fills forms, and extracts structured data from plain-English tasks
- PandaScript replay: agent sessions export to deterministic JavaScript that reruns without an LLM
- MCP server: stdio or HTTP transport, with isolated or shared browsing sessions per client
- Model providers: Anthropic, OpenAI, Gemini, Google Vertex AI, Mistral, Hugging Face, Vercel AI Gateway, OpenRouter, OpenAI-compatible endpoints, and local models via Ollama or llama.cpp
- Crawler controls: obeys robots.txt on request, with cookies, custom HTTP headers, proxy support, network interception, and an ad blocker
- Web Platform Tests: tested against the standardized WPT suite, with results published daily
