Teach your LLM agent to use Lightpanda, a fast headless browser optimized for data extraction and web automation, as a drop-in replacement for Chrome/Chromium.
This skill covers Lightpanda's interfaces and when to use them:
| Interface | Best for |
|---|---|
| MCP server | Agent workflows, interactive browsing, form filling |
| CLI fetch | Quick one-off page extraction |
| CDP server | Custom automation with Playwright/Puppeteer |
| Agent mode | Natural-language tasks, or authoring a PandaScript to save |
| Saved scripts (PandaScript) | Repeating a task deterministically, token-free |
A companion sub-skill, pandascript, covers writing PandaScript — the deterministic replay-script format run with lightpanda run (no LLM call).
SKILL.md and lightpanda.io/docs are the full reference — every tool, flag, and caveat.
Via skills.sh — works with Claude Code, Cursor, Codex, Windsurf, Copilot, opencode, OpenClaw, Hermes, and more:
npx skills add lightpanda-io/agent-skillIt prompts for the skills and target agent — pick lightpanda, plus pandascript if you'll write replay scripts.
Claude Code users can instead install the plugin, which also wires in the MCP server:
/plugin marketplace add lightpanda-io/agent-skill
/plugin install lightpanda@lightpandaWith the skill loaded, ask the agent to do it — it knows the install script, the package-manager options, and the platform quirks. Or run it yourself:
bash scripts/install.shHomebrew, AUR, and .deb builds: see the install docs.
The MCP server gives your agent live, interactive browsing — structured tools it calls mid-task (navigate, click, extract, wait, …) against a persistent session. It's one of the five interfaces above, not a requirement.
Claude Code plugin users already have it. For any other MCP client, add this to your config:
{ "mcpServers": { "lightpanda": { "command": "lightpanda", "args": ["mcp"] } } }- Linux (x86_64, aarch64)
- macOS (x86_64, arm64)
- Windows via WSL2
Apache 2.0