Lightpanda browser, drop-in replacement for Chrome-based browsing in any AI agent - faster and lighter for tasks without graphical rendering like data retrieval. Use it via MCP server, CLI fetch, or CDP with Playwright/Puppeteer — or run/save automations as deterministic, token-free replay scripts (

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: ${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](https://lightpanda.io/docs/run locally/installation/package managers): 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](https://github.com/lightpanda io/browser/releases) 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: Interface Best for How it works MCP server Agent workflows, interactive browsing, form filling Structured tools over stdio — purpose built for LLM agents CLI fetch Quick one off page extraction Single command, no server needed CDP server Custom automation with Playwright/Puppeteer WebSocket protocol, full browser control Agent mode One off natural language tasks, or authoring a PandaScript to save for later lightpanda agent — LLM driven CLI/REPL, optionally task "..." save script.js Saved scripts (PandaScript) Repeating the same task deterministically, without burning tokens Plain JS script, replayed with lightpanda run — no LLM call 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 To respect robots.txt , append obey robots to the command. Setup for other MCP clients Add to your MCP client configuration: 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: 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](https://lightpanda.io/docs/run locally/commands/fetch) is the complete reference. Examples Extract page as markdown: Extract semantic tree (compact, AI friendly): Fetch with longer wait for slow pages: CDP Server — Advanced Automation For full browser control via Playwright or Puppeteer: Start the Browser Server 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](https://lightpanda.io/docs/run locally/commands/serve) is the complete reference. Using with playwright core Connect using playwright core (not the full playwright package): 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. 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