The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The agent-browser install command downloads the Chrome/Chromium browser from a remote source.
[COMMAND_EXECUTION]: The skill utilizes a CLI for browser automation, supporting complex workflows through shell command chaining (&&) and batch execution.
[REMOTE_CODE_EXECUTION]: The eval command allows for the execution of arbitrary JavaScript within the browser context. The tool provides a Base64-encoded input option (-b) to facilitate passing scripts while avoiding shell escaping issues.
[DATA_EXFILTRATION]: Capabilities such as clipboard read, get text, and snapshot allow the agent to extract data from the system clipboard or web pages, which could include sensitive information.
[CREDENTIALS_UNSAFE]: The tool manages authentication states and session tokens. It includes a built-in 'Auth Vault' for encrypted storage and recommends using encryption keys (AGENT_BROWSER_ENCRYPTION_KEY) for protecting state files.
[PROMPT_INJECTION]: The skill documentation addresses indirect prompt injection risks from untrusted web content and provides an opt-in 'Content Boundaries' feature to mitigate them.
Ingestion points: Web content is ingested into the agent context via snapshot, get text, and screenshot --annotate.
Boundary markers: The AGENT_BROWSER_CONTENT_BOUNDARIES feature uses nonce-based delimiters to separate tool output from page content.
Capability inventory: The agent can navigate, interact with forms, execute scripts, manage files, and handle browser sessions.
Sanitization: Content boundaries provide a structural mechanism to help the AI distinguish and ignore malicious instructions embedded in web pages.

agent-browser