The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes several local shell commands for its core functionality. It uses osascript to query Google Chrome tabs for instance identification, ps and pgrep to identify the parent application of the ChatGPT extension, and open to launch applications or URLs on the host system. The installation script uses jq to modify local configuration files.
[PERSISTENCE]: The install-codex-cu.sh script registers the MCP server at the user scope by modifying configuration files for Claude Code (~/.claude.json) and Cursor (~/.cursor/mcp.json). This ensures the tool remains available across agent sessions.
[INDIRECT_PROMPT_INJECTION]: As a Computer Use bridge, the skill ingests untrusted data from application accessibility trees and browser DOM snapshots. Ingestion points include getAXState() for native apps and domSnapshot() for web pages. There are no explicit boundary markers enforced by the proxy scripts. The capability inventory is high, including application control, browser manipulation, and JavaScript execution. Sanitization is not present, as the skill acts as a transparent bridge.
[DYNAMIC_EXECUTION]: The skill facilitates the execution of arbitrary JavaScript within target browser tabs and provides high-level UI automation through the cua_repl MCP server. It also uses os.execvpe to launch the underlying plugin binary dynamically.

codex-cu