The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute arbitrary shell commands using exec and zsh -lc, granting broad access to the local operating system and user environment.
[REMOTE_CODE_EXECUTION]: The skill provides instructions for dynamic execution of Python code via browser-use python and JavaScript via browser-use eval. These capabilities allow for the execution of complex, runtime-generated logic that can interact with the system or browser state.
[EXTERNAL_DOWNLOADS]: The skill authorizes the installation of external packages and CLI tools using uv tool install, bun add -g, and brew install. While these are standard package managers, the ability for an agent to install unverifiable third-party code from the internet poses a significant supply chain risk.
[DATA_EXFILTRATION]: The skill includes explicit commands for exporting browser cookies (browser-use cookies export) and capturing screenshots. These tools could be leveraged to harvest sensitive session data or private information from the user's browser.
[PROMPT_INJECTION]: The skill processes untrusted data from external URLs (web_fetch) and local documents (markitdown). It lacks robust boundary markers or explicit instructions to ignore embedded commands, making it vulnerable to indirect prompt injection where malicious content in a document or webpage could hijack the agent's behavior.

local-tools-guide