The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill documentation mentions the agent-browser install command for environment setup. The tool is hosted by Vercel Labs, which is a trusted and well-known technology organization.
[COMMAND_EXECUTION]: Commands are executed via the Bash tool to control the agent-browser CLI. This includes an eval feature for running JavaScript in the browser context, which is a core automation function but necessitates caution when handling dynamic inputs.
[PROMPT_INJECTION]: The skill is subject to indirect prompt injection as it processes content from external websites. \n
Ingestion points: Loads and snapshots web pages via agent-browser open and snapshot as defined in SKILL.md and templates/capture-workflow.sh. \n
Boundary markers: Recommends using AGENT_BROWSER_CONTENT_BOUNDARIES to wrap page data in delimiters, reducing the likelihood of the agent obeying embedded instructions. \n
Capability inventory: The skill uses Bash, Write, and Read tools, enabling it to perform shell operations and persist data locally. \n
Sanitization: Provides hardening via AGENT_BROWSER_ALLOWED_DOMAINS and AGENT_BROWSER_ACTION_POLICY to restrict the agent's reach and available actions.
[DATA_EXFILTRATION]: Supports saving browser session states (cookies and auth tokens) to local files like auth.json (references/authentication.md). While this allows for authentication persistence, these files contain sensitive data and should be protected from unauthorized access.

agent-browser