The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the agent-browser CLI tool for browser control. It includes instructions for complex command sequences to automate web interactions, navigation, and local file access using the --allow-file-access flag.\n- [REMOTE_CODE_EXECUTION]: The agent-browser eval command allows the agent to execute arbitrary JavaScript within the browser context. Documentation includes best practices for using Base64 encoding and stdin to prevent shell-related command corruption.\n- [DATA_EXFILTRATION]: The skill provides capabilities to extract page content, take screenshots, and save entire browser session states (including cookies and localStorage) to the local filesystem. While these are core automation features, they involve handling potentially sensitive user data.\n- [CREDENTIALS_UNSAFE]: The skill manages authentication using environment variables and browser state files. It includes clear security warnings regarding the sensitive nature of these files and advises against committing them to version control.\n- [EXTERNAL_DOWNLOADS]: The skill utilizes the agent-browser package from the NPM registry and references the Appium framework for mobile automation tasks.\n- [PROMPT_INJECTION]: The skill ingests untrusted data from external websites through commands like snapshot and get text.\n
Ingestion points: agent-browser snapshot, agent-browser get text (SKILL.md, references/snapshot-refs.md)\n
Boundary markers: None identified in the prompt templates.\n
Capability inventory: File writes (state save, screenshot, pdf), command execution (agent-browser), and JavaScript execution (eval).\n
Sanitization: No explicit sanitization or validation of the ingested web content before processing by the agent.

agent-browser