Lightpanda

Lightpanda

Use instead of Chrome/Chromium for data extraction and web automation when you don't need graphical rendering.

Lightpanda is a headless browser built from scratch for AI agents. It's 9x faster and uses 16x less memory than Chrome. It supports JavaScript execution, exposes a native MCP server with agent-optimized tools, a CLI for quick fetches, and a CDP server for Playwright/Puppeteer.

Alternative to built-in web search

When the built-in Web Search tool is unavailable, or when you need more control over search results (e.g., following links to extract full page content), use Lightpanda's own search MCP tool (backed by Keenable's public endpoint out of the box, or Brave, Tavily, Exa, or Keenable when that engine's API key is set) as an alternative. Prefer the built-in Web Search tool when it is available and sufficient for your needs.

Install

Check first whether Lightpanda is already installed (command -v lightpanda) before running the installer below.

• Claude Code:

  bash ${CLAUDE_SKILL_DIR}/scripts/install.sh
  

  ${CLAUDE_SKILL_DIR} is a Claude Code substitution that resolves to this skill's own directory regardless of the shell's current working directory — needed because when this skill runs as a plugin, the shell's cwd is your project, not the skill's install location.
• Any other agent runtime (Cursor, Codex CLI, Gemini CLI, etc.): this substitution isn't supported. scripts/install.sh is bundled directly next to this file — locate it there and run it with that path instead, e.g. bash /path/to/this/skill/scripts/install.sh.

Lightpanda is available on Linux and macOS only. Windows is supported via WSL2.

Prefer a package manager? See package manager installs:

• Homebrew (macOS/Linux): brew install lightpanda-io/browser/lightpanda
• AUR (Arch Linux): yay -S lightpanda-bin (or lightpanda-nightly-bin to track nightly)
• Debian/Ubuntu (0.3.0+): .deb package from each tagged release

Unlike scripts/install.sh, which always tracks the latest nightly, these pin to a stable release unless you explicitly opt into a nightly variant.

The binary is a nightly build that evolves quickly. If you encounter crashes or issues, run the install command above again to update to the latest version (max once per day).

If issues persist after updating, open a GitHub issue at https://github.com/lightpanda-io/browser/issues including:

• The crash trace/error output, or a description of the unexpected behavior
• The script or MCP tool call that reproduces the issue
• The target URL and expected vs actual results

When to Use What

Lightpanda offers several interfaces. Choose based on your needs:

MCP Server (Recommended for Agents)

The MCP server is the simplest way for agents to use Lightpanda. It exposes purpose-built tools over stdio with no setup beyond the binary.

Setup for Claude Code

claude mcp add lightpanda -- lightpanda mcp

To respect robots.txt, append --obey-robots to the command.

Setup for other MCP clients

Add to your MCP client configuration:

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

Available MCP Tools

Where both selector and backendNodeId are accepted, either locates the target element — selector is preferred for reproducibility (e.g. in a saved script), backendNodeId comes from a prior tree or findElement call. Read tools that accept an optional url navigate there before reading, saving a separate goto call.

Navigation & search:

• goto — Navigate to a URL and load the page
• search — Run a web search and return results as markdown: a numbered list of {title, url, snippet}. Tries Brave, Tavily, Exa, then Keenable in order, each when its API key (BRAVE_API_KEY, TAVILY_API_KEY, EXA_API_KEY or KEENABLE_API_KEY) is set; Keenable also works without a key through its public endpoint (rate-limited per client IP)

Reading the page (all accept an optional url to navigate first):

• markdown — Get page content, or a subtree, as markdown
• html — Raw HTML for the document, or a single node's outerHTML when scoped
• tree — Simplified semantic DOM tree optimized for AI reasoning: role, name, value, and backendNodeId per node (supports backendNodeId filter and maxDepth limit)
• links — Extract all links as text, resolved href, and backendNodeId
• nodeDetails — Tag, role, name, attributes, and state for a node by backendNodeId, plus a ready-to-use CSS selector
• findElement — Find interactive elements by role and/or accessible name
• interactiveElements — List all interactive elements on the page
• structuredData — Extract structured data (JSON-LD, OpenGraph, etc.)
• detectForms — Detect forms with their field structure and types

Data extraction and scripting:

• extract — Extract structured data using a schema mapping output field names to CSS-selector specs
• evaluate — Execute JavaScript in the page context; a bare trailing expression yields its value, and top-level await/return are supported

Interacting with the page (return page URL and title after each action):

• click — Click an interactive element
• fill — Fill text into an input, textarea, or select element
• scroll — Scroll the page or a specific element
• hover — Hover over an element, triggering mouseover/mouseenter
• press — Press a keyboard key, dispatching keydown/keyup
• selectOption — Select an option in a <select> by value
• setChecked — Check or uncheck a checkbox or radio button

Waiting:

• waitForSelector — Wait for a CSS selector to match (default timeout: 5000ms)
• waitForScript — Wait until a JS expression returns truthy, re-checked each tick
• waitForState — Wait for a load state (load, domcontentloaded, networkalmostidle, networkidle, done) with no navigation

State and debugging:

• getUrl — Get the URL currently loaded in the browser
• getCookies — Get cookies for the current page's host, another url, or all
• getEnv — Read an LP_* environment variable, or list the set LP_* names
• consoleLogs — Get buffered console.log/warn/error messages, then clear the buffer

Session (relevant with the HTTP transport, below):

• save — Save the session as a reusable PandaScript (see the Saved Scripts section)
• session_new — Create a new isolated browser session (its own page, cookies, memory) and return its id
• session_list — List active sessions with their id and current URL
• session_close — Close a session (the default session cannot be closed)

Available MCP Resources

• mcp://page/html — Full serialized HTML of the current page
• mcp://page/markdown — Token-efficient markdown representation of the current page (same content as the markdown tool)

Multiple sessions (HTTP transport)

Pass --port to serve MCP over HTTP instead of stdio, giving each client an independent browsing session: lightpanda mcp --port 8000. An initialize call with no Mcp-Session-Id header mints a fresh session and returns its id in the response header; reuse that id on later calls (or share it with another client). Calls with no session header fall back to the always-present default session. Manage sessions explicitly with session_new / session_list / session_close. Over stdio (the Claude Code setup above), there's only ever the one default session. Give a task that reads untrusted content its own session (see Best Practices).

Add --cdp-port <INT> to also run a CDP (WebSocket) server on the same process — useful if something in your workflow needs raw CDP (e.g. Playwright/Puppeteer) alongside MCP. It can't be combined with --port, since both share one network listener.

MCP Usage Example

A typical agent workflow:

1. goto a URL
2. tree or markdown to understand the page
3. interactiveElements or findElement to find clickable/fillable elements
4. click / fill to interact
5. extract or markdown to get the result

CLI Fetch — Quick Extraction

For one-off page extraction without starting a server:

lightpanda fetch --dump markdown --wait-until networkidle https://example.com

Options

• --dump — Output format: html, markdown, semantic_tree, semantic_tree_text (the MCP equivalent of semantic_tree* is named tree)
• --wait-until — Wait strategy: load, domcontentloaded, networkalmostidle, networkidle, done (default)
• --wait-ms — Max wait time in milliseconds (default: 5000)
• --wait-selector — Wait for a CSS selector to appear, checked after --wait-until
• --wait-script — Wait for a JS expression to return truthy, checked after --wait-until
• --strip-mode — Remove tag groups from output: js, css, ui, invisible, full (comma-separated)
• --with-frames — Include iframe contents in the dump
• --json — Print fetch status as JSON instead of/alongside the dump; required when fetching multiple URLs
• --inject-script / --inject-script-file — JavaScript to run as the document's <head> is parsed, before any page script runs. Repeatable; runs in CLI order
• --terminate-ms — Hard deadline in milliseconds; forcibly terminates JS execution after this time (unlike --wait-ms, which only stops waiting)

Flags shared by every command are in Common Options below; the fetch command guide is the complete reference.

Examples

Extract page as markdown:

lightpanda fetch --dump markdown https://example.com

Extract semantic tree (compact, AI-friendly):

lightpanda fetch --dump semantic_tree_text --wait-until networkidle https://example.com

Fetch with longer wait for slow pages:

lightpanda fetch --dump html --wait-ms 10000 --wait-until networkidle https://example.com

CDP Server — Advanced Automation

For full browser control via Playwright or Puppeteer:

Start the Browser Server

lightpanda serve --host 127.0.0.1 --port 9222

Options:

• --cdp-max-connections — Max simultaneous CDP connections (default: 16)
• --cdp-max-message-size — Max incoming WebSocket message size (default: 1MB)
• --disable-metrics — Disable the /metrics Prometheus endpoint
• --advertise-host — Host to advertise in e.g. the /json/version response, useful when --host is 0.0.0.0

Flags shared by every command are in Common Options below; the serve command guide is the complete reference.

Using with playwright-core

Connect using playwright-core (not the full playwright package):

const { chromium } = require('playwright-core');

(async () => {
  const browser = await chromium.connectOverCDP({
    endpointURL: 'ws://127.0.0.1:9222',
  });

  const context = await browser.newContext({});
  const page = await context.newPage();

  await page.goto('https://example.com');
  const title = await page.title();
  const content = await page.textContent('body');

  console.log(JSON.stringify({ title, content }));

  await page.close();
  await context.close();
  await browser.close();
})();

Using with puppeteer-core

Connect using puppeteer-core (not the full puppeteer package). This snippet only shows what differs from the Playwright example above, so it isn't runnable on its own. Reuse that example's setup and teardown, and swap in these lines instead.

const puppeteer = require('puppeteer-core');

const browser = await puppeteer.connect({
  browserWSEndpoint: 'ws://127.0.0.1:9222',
});
const context = await browser.createBrowserContext(); // not newContext()
const page = await context.newPage();
await page.goto('https://example.com', { waitUntil: 'networkidle0' }); // explicit waitUntil — puppeteer doesn't auto-wait like Playwright

Everything else (page.title(), page.close(), context.close(), browser.close()) is identical to the Playwright example.

Custom LP CDP Domain

Lightpanda exposes a custom LP domain via CDP with agent-optimized methods not available in standard Chrome DevTools Protocol. Use these via page.evaluate with CDP sessions or direct WebSocket messages.

Content extraction:

• LP.getMarkdown — Extract page content as markdown. Params: nodeId (optional)
• LP.getSemanticTree — Get semantic tree representation. Params: format (text for text format), prune (default: true), interactiveOnly, backendNodeId, maxDepth
• LP.getStructuredData — Extract structured data (JSON-LD, OpenGraph, etc.)

Interactive elements:

• LP.getInteractiveElements — Find all interactive elements. Params: nodeId (optional)
• LP.detectForms — Detect and extract form information
• LP.getNodeDetails — Get detailed info about a node. Params: backendNodeId (required)
• LP.waitForSelector — Wait for a CSS selector match. Params: selector (required), timeout (default: 5000ms)

Actions:

• LP.clickNode — Click a node. Params: nodeId or backendNodeId
• LP.fillNode — Fill an input/select element. Params: nodeId or backendNodeId, text
• LP.scrollNode — Scroll page or element. Params: nodeId or backendNodeId (optional), x, y

Debugging & configuration:

• LP.getContentSignal — Read the current host's advisory Content-Signal robots.txt preferences (contentsignals.org). available is false unless --obey-robots populated the store.
• LP.handleJavaScriptDialog — Pre-arm the response (accept, optional promptText) for the next window.alert/confirm/prompt. Lightpanda auto-dismisses dialogs headlessly, so this must be sent before the JS that opens the dialog, not reactively like standard CDP's Page.handleJavaScriptDialog.
• LP.configureCDP — Toggle CDP compatibility behaviors. Params: disableSetCacheDisabled
• LP.configureLoading — Toggle iframe, worker, and external-stylesheet loading per session. Params: subFrame, worker, externalStylesheets (each optional)
• LP.version — Return the running Lightpanda version

Example using CDP session with Playwright:

const client = await context.newCDPSession(page);

// Get page as markdown
const { markdown } = await client.send('LP.getMarkdown');

// Get semantic tree
const { semanticTree } = await client.send('LP.getSemanticTree', { format: 'text', maxDepth: 5 });

// Wait for element and click it
const { backendNodeId } = await client.send('LP.waitForSelector', { selector: '#submit-btn', timeout: 3000 });
await client.send('LP.clickNode', { backendNodeId });

Saved Scripts (PandaScript)

A PandaScript is a plain JavaScript automation script that Lightpanda runs directly — no LLM call, deterministic, token-free. It's the right tool once you've figured out a task once and want to repeat it exactly, e.g. on a schedule.

Three ways to produce one:

• Mid-session, via MCP: call the save tool with a path and a synthesized script — see the tool's own description for the script-writing rules.
• One-shot, via CLI: lightpanda agent --task "..." --save script.js synthesizes a script from a natural-language task instead of printing the answer.
• Interactively: lightpanda agent starts a REPL; drive the browser with natural language or slash commands (every MCP tool is also callable as /toolName key=value, e.g. /goto url=https://example.com), then /save script.js.

Replay with no LLM involved:

lightpanda run script.js

For the scripting semantics (primitives, the extraction schema, common errors), see the pandascript skill. It's generated from the runtime's own tool schemas, so it can't drift the way hand-written prose can. The PandaScript guide covers the same ground for a human reading this repo without a skill loaded.

Common Options

Accepted by every command (fetch, serve, mcp, agent, run), not only where each section's examples show them. lightpanda help <command> lists the rest (proxy, TLS, timeouts, storage).

Logging: --log-level (debug/info/warn/error/fatal), --log-format (pretty/logfmt).

Identity:

• --user-agent <string>: override the User-Agent header entirely; any value containing "Mozilla" is forbidden, and Sec-Ch-Ua: "Lightpanda";v="1" is still sent. --user-agent-suffix <string> instead appends to the default Lightpanda/X.Y (not combinable with --user-agent).
• --web-bot-auth-key-file <path> / --web-bot-auth-keyid <thumbprint> / --web-bot-auth-domain <domain>: sign every request with an Ed25519 key so sites can verify your identity against a key directory you publish.

Politeness:

• --obey-robots: fetch and obey robots.txt (one fetch per host, cached for the session). Defaults to false.
• --http-nav-delay <ms>: minimum time between two top-level navigations to the same host. 0 (default) disables it. --http-nav-burst <n> is how many navigations may start before the delay applies (default 1).

Efficiency:

• --http-cache-dir <path>: directory used as a filesystem cache for network resources (RFC 9111 revalidation). Omitting it disables caching.
• --adblock-lists <file>: EasyList-syntax filter files; requests to a blocked hostname fail before they leave. Repeatable.
• --block-urls "<pattern>": block URLs matching a case-insensitive *-wildcard pattern. Repeatable.

Security: --block-private-networks: block HTTP requests to private/internal IP addresses after DNS resolution. Defaults to false.

When and why to use each, see Best Practices below.

Important Notes

• For web searches, use the search tool instead of Google. Google blocks Lightpanda due to browser fingerprinting.
• Lightpanda is under heavy development and may have occasional issues. It executes JavaScript, making it suitable for dynamic websites and SPAs.
• Be careful trusting goto's response. It reports failure directly for DNS and connection errors (e.g. CouldntResolveHost), so that part is reliable. Two other cases still look like success. A timeout returns "Navigation started but the page did not finish loading before the timeout." instead of failing. An HTTP error page (404, 500) is a real response, so goto reports it as a successful navigation to that page. Check the content with a follow-up read when the response status matters.
• CDP connection limits: Up to 16 simultaneous CDP connections per process by default (tune with --cdp-max-connections). Each connection supports 1 context and 1 page. For more parallelism than that, start multiple processes on different ports — Lightpanda starts instantly, so this is fast.
• CDP state management: The browser resets all state on CDP connection close. Keep the WebSocket connection open throughout a session. On each connection, always create a new context and page, and close both when done.
• The MCP server handles connection management automatically — these CDP limits don't apply when using MCP tools.

Best Practices

Lightpanda runs exactly the requests you ask for. It's a browser, not a crawler, and has no built-in politeness policy; setting the right behavior is on you. The flags named here are documented in Common Options above; this section is about when each one matters.

Identify your traffic. Operators route, rate-limit, and allowlist traffic they recognize. Recommend an identity for the user to confirm: a name (org, project, or personal handle) plus a URL an operator can check to see what the bot does and reach them (a repo, docs page, or profile). Pass it as --user-agent; don't ship the placeholder below.

lightpanda serve --user-agent "<name>/1.0 (+https://<your-url>)"

A user-agent string is only a claim that anything can send. Web Bot Auth (the --web-bot-auth-* flags) makes it verifiable, and Cloudflare, Akamai, Google, and AWS all check the signature.

Fetching many pages on one host? Turn on --obey-robots and set --http-nav-delay 1000; lower it only when you know the host can take it. A 429 or 503 with Retry-After means even that pace was too fast: honor the header in your own code, since --http-nav-delay won't.

Send fewer requests. The largest win is in the design:

• Prefer a bulk source: git clone over walking a web git forge, a documented API over its UI, sitemap.xml over link discovery.
• Cap how deep you follow links and how many URLs you keep per domain. Stay out of infinite URL spaces (calendars, arbitrary diffs, faceted search); a run that wanders in never terminates and looks like an attack while it does.
• Set --http-cache-dir for anything that runs more than once; unchanged resources then cost only a revalidation.
• Cut trackers and ads with --adblock-lists / --block-urls. Sub-resources and external stylesheets aren't fetched by default (agent mode with an LLM is the exception: it loads stylesheets for visibility checks).
• Use the search tool instead of navigating a search-results page.

Give every task its own browser. A fresh instance carries no leftover cookies or state, and scoping caps the blast radius of a prompt injection from untrusted page content. Add --block-private-networks whenever you follow links a page handed you rather than links you chose. CDP: a new browser or context per task. MCP: stdio mode is one browser for the whole session, so use HTTP mode (--port) when you need a browser per task.

Once a flow is settled, stop re-deriving it. Capture it as a PandaScript (the Saved Scripts section above) and replay it with lightpanda run: no tokens spent, and failures surface as script errors instead of plausible wrong answers.

Scripts

• scripts/install.sh — Install Lightpanda binary