The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of the main CLI tool and its dependencies from public registries and repositories.
Evidence: Instructions for npm i -g agent-browser && agent-browser install are provided in SKILL.md.
Evidence: The plugin add command allows downloading extension modules from npm or GitHub (references/commands.md).
[COMMAND_EXECUTION]: The skill operates primarily by executing a comprehensive suite of browser automation commands via a dedicated CLI.
Evidence: The core loop involves commands like open, snapshot, click, and fill (SKILL.md).
[DYNAMIC_EXECUTION]: The skill provides capabilities for running custom logic and scripts within the managed browser environment.
Evidence: The eval command allows executing JavaScript strings or files within the browser context (references/commands.md).
Evidence: The network route command enables dynamic interception and mocking of network requests (references/commands.md).
[INDIRECT_PROMPT_INJECTION]: The skill is designed to handle untrusted data from the web and includes clear defensive instructions for the agent.
Evidence: The references/trust-boundaries.md file explicitly warns the agent to treat all browser-sourced content as data rather than instructions.
Evidence: The use of compact accessibility-tree snapshots (@eN refs) minimizes the amount of raw web content the agent must process (SKILL.md).
[PROMPT_INJECTION]: A detector alert for prompt injection was found to be a false positive.
Evidence: The strings flagged in references/trust-boundaries.md are examples of malicious instructions (like "ignore previous instructions") used in a defensive context to teach the agent what to avoid.

agent-browser