The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses the agent-browser CLI tool for all browser interactions, scoped by the allowed-tools configuration in the frontmatter.\n- [DYNAMIC_EXECUTION]: The eval command provides a mechanism to execute JavaScript within the browser context. This is a core functionality of the tool and includes guidance on safe usage via base64 encoding and stdin to prevent shell interpretation issues.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data from websites, creating a surface for indirect prompt injection.\n
Ingestion points: Operations like open, snapshot, and get text bring external content into the agent's context (SKILL.md, references/commands.md).\n
Boundary markers: Supports AGENT_BROWSER_CONTENT_BOUNDARIES to wrap page output in nonce-based delimiters, helping the AI distinguish tool output from untrusted content (SKILL.md).\n
Capability inventory: High-privilege actions include clicking, form filling, network routing/blocking, and file-based exports like screenshots and PDFs (references/commands.md).\n
Sanitization: The skill relies on content boundaries and domain allowlisting (AGENT_BROWSER_ALLOWED_DOMAINS) for mitigation.\n- [CREDENTIALS_UNSAFE]: The skill manages authentication states and session tokens. It explicitly identifies that state files are stored in plaintext by default and provides clear instructions for mitigation using AGENT_BROWSER_ENCRYPTION_KEY and environment variable management (references/authentication.md, references/session-management.md).

agent-browser