The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The scripts/install.sh script downloads the Lightpanda binary from the official Lightpanda-io GitHub repository (github.com/lightpanda-io/browser/releases/download/nightly/). The script performs integrity verification by fetching the SHA256 digest from the GitHub API and validating the downloaded file before granting execution permissions.
[COMMAND_EXECUTION]: The skill requires execution of the lightpanda binary and its installation script. These capabilities are explicitly scoped in the allowed-tools metadata in SKILL.md to ensure the agent has the necessary permissions to operate the browser.
[INDIRECT_PROMPT_INJECTION]: The skill exposes a significant attack surface for indirect prompt injection as it ingests untrusted data from the web.
- Ingestion points: Tools such as goto, markdown, html, tree, and search (in SKILL.md) fetch content from arbitrary external URLs.
- Boundary markers: The documentation recommends using --obey-robots and session isolation via --port in HTTP mode to limit cross-task impact.
- Capability inventory: The browser can perform interactive actions (click, fill, evaluate), save sessions (save tool), and make outbound network requests.
- Sanitization: Not explicitly handled within the skill scripts; relies on the agent's platform-level guardrails when processing tool outputs.
[DYNAMIC_EXECUTION]: The skill supports dynamic JavaScript execution through PandaScripts (lightpanda run) and the page.evaluate MCP tool. This is a core feature of the headless browser designed for dynamic web automation.
[CREDENTIALS_SAFE]: The skill follows security best practices by using environment variables (LP_*) and placeholders ($LP_*) for managing API keys and passwords, avoiding hardcoded secrets in scripts or instructions.

lightpanda