Lightpanda Browser
A lightweight, scriptable headless browser for AI and automation workflows.
- Source repo
- lightpanda-io/browser
- Stars
- ★ 37k
- Last updated
- today
- License
- AGPL-3.0
- Primary language
- Zig
- FA score
- 66/100 · Some gaps
At a glance
- How it runs
- Works with
- Universal · cross-platformOpenAI API · Claude API
- Cost
- Free, no paid service needed
- Setup effort
- Medium · a few setup steps
- You'll need
- Typical use
- Data collection developers who need to fetch many pages and save them as HTML or Markdown.
- Not a fit if
- Teams that require a native Windows browser
- Workflows that depend on a full graphical rendering engine
- Source review
- 66/100 · Some gaps
What does this agent do, and when should you use it?
Lightpanda is a headless browser written from scratch in Zig; it is not based on Chromium, Blink, or WebKit and has no graphical rendering engine. Its CLI fetches pages and exports content, while its browser server exposes CDP and WebDriver Bidi for clients such as Puppeteer. The built-in lightpanda agent can operate the browser from natural-language tasks and save sessions as replayable JavaScript PandaScript. Lightpanda also offers an MCP server over stdio or HTTP and an optional Docker deployment. In the README's stated AWS EC2 m5.large benchmark, it reports 123 MB peak memory and 5 seconds for 100 pages; those figures apply to that specific test setup.
Run lightpanda fetch to load a URL and export HTML, Markdown, PNG, PDF, or structured metadata. Run lightpanda serve to expose a CDP service, optionally with WebDriver Bidi, so clients such as Puppeteer can navigate pages, execute JavaScript, inspect the DOM, click, and fill forms. lightpanda agent accepts tasks through a configured model provider, operates pages, and extracts structured data; /save exports a PandaScript that can be replayed with lightpanda run. Run lightpanda mcp to provide browser tools to MCP clients over stdio or HTTP.
- Data collection developers who need to fetch many pages and save them as HTML or Markdown.
- Automation teams that want Puppeteer or Playwright to navigate pages, click controls, and fill forms.
- Content teams that need to extract JSON-LD, OpenGraph, Twitter card, or other page metadata.
- Engineers who want to prototype browser tasks in natural language, then save and replay them as JavaScript.
- Self-hosting teams that want to provide browser sessions to one or more agents through MCP.
How do you install or deploy this agent?
Download a nightly binary for Linux or macOS, or run the Docker image. Linux release binaries require glibc; Windows users need to install Lightpanda inside WSL2. Source builds require Zig 0.15.2 and the dependencies listed in the README.
brew install lightpanda-io/browser/lightpandayay -S lightpanda-nightly-bindocker run -d --name lightpanda -p 127.0.0.1:9222:9222 lightpanda/browser:latestHow do you use this agent?
Fetch a page and export Markdown:
./lightpanda fetch --obey-robots --dump markdown https://example.comStart a CDP server:
./lightpanda serve --host 127.0.0.1 --port 9222Start the MCP stdio server and add this configuration to your MCP client:
{
"mcpServers": {
"lightpanda": {
"command": "/path/to/lightpanda",
"args": ["mcp"]
}
}
}For the built-in agent, API keys can be detected from the environment; a no-model REPL is also available. Model-driven tasks require credentials for a supported provider.
./lightpanda agent --task "top story on news.ycombinator.com?"
./lightpanda agent --no-llm
./lightpanda run session.jsWhat are this agent's strengths and limitations?
- Built from scratch in Zig without Chromium, Blink, or WebKit; the README reports about 16x lower peak memory and 9x faster execution than Headless Chrome in its stated benchmark.
- Supports CDP and WebDriver Bidi, with documented connections from browser automation clients such as Puppeteer.
- Offers page exports, a built-in agent, replayable PandaScript, and an MCP server.
- The agent documentation lists Anthropic, OpenAI, Gemini, Vertex AI, Mistral, Hugging Face, Vercel AI Gateway, OpenRouter, OpenAI-compatible endpoints, and local models.
- There is no native Windows binary; Windows users need WSL2.
- Linux release binaries are linked against glibc and do not run directly on musl distributions such as Alpine.
- Model-driven agent tasks require provider credentials, and the README does not state model call prices.
- There is no graphical rendering engine, so workflows that depend on full visual rendering may not fit.
How does this agent compare with similar options?
The README benchmarks Lightpanda against Headless Chrome for memory and execution time and clarifies that Lightpanda is not a Chromium fork. It also contrasts building a browser from scratch with forking Chromium.
Key facts side by side with the most closely related agents.
| Agent | Source review | Form / cost | Stars | Updated | Language | Full support on |
|---|---|---|---|---|---|---|
| Lightpanda Browser This agent | 66 · Some gaps | CLIFree | ★ 37k | today | Zig | OpenAI API · Claude API |
| Obscura Headless Browser | 75 · Good | CLIFree | ★ 29k | today | Rust | Claude.ai |
| Flyto2 Core | 64 · Some gaps | CLIFreemium | ★ 484 | 8d ago | Python | Claude Code |
| Invisible Browser Agent | 57 · Major gaps | MCP serverFree + model costs | ★ 1.3k | 3d ago | Python | Codex · Claude Code |
How does FollowAgents rate this agent?
Why each dimension lost points
The README lists browser navigation, clicking, form filling, data extraction, CDP/MCP servers, proxy support, and network interception. It also says usage telemetry is collected and sent by default and can be disabled with an environment variable, but does not specify telemetry fields or transport details. The materials do not show sensitive-credential handling, permission scoping, or confirmation before side effects. Localhost binding and a mode-0600 Unix socket offer some access-control guidance, but no clear operation rollback is described. A private security reporting channel is specified, and workflow Actions use pinned commit references; dependency security review and verified publisher attribution remain insufficiently documented.
The README separates build, installation, server startup, tests, and browser capabilities, and the main usage paths are coherent. It specifies a Zig version and build dependencies and provides unit, end-to-end, and WPT procedures. It documents musl, Android/Termux, and native Windows limitations with failure symptoms or alternatives. No code or tests were run; runtime behavior and test reproducibility are outside this static review, so the evidence does not justify full marks.
The documentation clearly targets AI agents and automation, with examples for Puppeteer, Playwright, MCP, CDP, WebDriver Bidi, proxies, multiple model providers, and an LLM-free REPL. Explicit commands start agent, fetch, serve, or MCP modes, making activation reasonably clear, though the agent's action boundaries are not fully described. Linux, macOS, WSL, Docker, glibc, and architecture differences receive concrete environment guidance.
The README has clear sections, installation options, command examples, integration examples, build and test instructions, and several platform limitations. The LICENSE file contains the AGPL-3.0 text. Some version information, including Zig 0.15.2, is explicit, but the supplied materials do not include a changelog or release policy. The security reporting address and supported-version policy are clear; routine maintenance ownership and verified identity are not.
The product offers HTML, Markdown, PNG, PDF, and structured-data output, a native browser agent, MCP, and saveable, replayable PandaScript, all of which support downstream automation. The benchmark table gives page count, instance type, memory, and timing comparisons, but these remain repository claims and were not independently verified here. Specific performance figures support a potential speed and resource advantage, while total cost and suitability tradeoffs are less developed.
Installation commands, environment variables, protocols, test commands, benchmark conditions, and several feature claims have concrete text or examples. The README points to external benchmark and WPT results, and a workflow shows scheduled agent regression with pinned Action versions. This provides some cross-source support, but the external pages are not included in the supplied evidence, so performance and test claims cannot be independently checked in this review. The documentation distinguishes implemented features from some platform limitations, but gives less detail on validation boundaries for benchmark claims and on agent behavior.
- Telemetry is enabled by default; check your organization's privacy requirements and review the privacy policy and telemetry data details before use.
- The tool can access websites, fill forms, and run browser automation. Restrict network exposure when deploying its servers and handle model API keys, cookies, and page content carefully.
- Performance figures and linked test results in the README were not independently verified in this static review.
FAQ
Do I need an LLM API key to use it?
lightpanda agent --no-llm starts a REPL without an LLM. Model-driven tasks require credentials for the chosen provider.