The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The documentation instructs users to install the agent-browser package from the NPM registry. It recommends pinning specific versions and installing in dedicated environments to mitigate supply chain risks.
[COMMAND_EXECUTION]: The skill utilizes the Bash tool to run CLI commands for web automation. It provides guidance on safe command sequences and operational guardrails for AI agents.
[REMOTE_CODE_EXECUTION]: The documentation describes the eval command, which allows for arbitrary JavaScript execution within a browser session. This is correctly identified as a high-risk capability that should only be used with explicit human approval.
[PROMPT_INJECTION]: As a tool designed for web automation and snapshotting, the skill creates an attack surface for indirect prompt injection from untrusted web content.
Ingestion points: Target URLs loaded via open and element content captured via snapshot (SKILL.md, agent-browser-workflows.md).
Boundary markers: The documentation suggests using structured snapshots and JSON output for deterministic parsing but does not define explicit delimiters for untrusted data.
Capability inventory: The agent has full browser control (click, fill, select) and can perform filesystem writes via the download command (references/agent-browser-command-map.md).
Sanitization: The skill explicitly recommends domain allowlisting, redacting tokens in logs, and avoiding file access unless required (references/agent-browser-safety.md).

agent-browser-core