The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes a set of Bash commands under the agent-browser namespace to perform browser automation tasks such as open, click, fill, and screenshot.
[EXTERNAL_DOWNLOADS]: The skill uses npx to execute the agent-browser package, which involves fetching the tool from the npm registry at runtime.
[DYNAMIC_EXECUTION]: Provides an eval command that allows the agent to execute arbitrary JavaScript within the browser context for complex data extraction or interaction tasks.
[INDIRECT_PROMPT_INJECTION]: As a tool designed to process web content, it is inherently exposed to indirect prompt injection from malicious websites. The skill documentation explicitly recommends using the --content-boundaries flag to help the agent distinguish between tool instructions and untrusted page content.
Ingestion points: agent-browser snapshot, agent-browser get text, and agent-browser screenshot fetch data from external URLs into the agent's context.
Boundary markers: Supports opt-in markers via the AGENT_BROWSER_CONTENT_BOUNDARIES environment variable.
Capability inventory: The tool can navigate to any URL, interact with forms, and execute JavaScript, controlled via Bash tool calls.
Sanitization: The skill provides AGENT_BROWSER_ALLOWED_DOMAINS to restrict navigation to trusted sources and an AGENT_BROWSER_ACTION_POLICY to gate destructive actions.
[DATA_EXFILTRATION]: The skill manages sensitive authentication states and session cookies. It mitigates exposure risks through an 'Auth Vault' that encrypts credentials using a user-provided AGENT_BROWSER_ENCRYPTION_KEY. It also supports local file access via the --allow-file-access flag, which is documented for legitimate local document processing.

agent-browser