The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a command-line interface (open-computer-use and its alias ocu) to perform desktop actions, list applications, and manage system-level configurations.
[EXTERNAL_DOWNLOADS]: The skill instructs the agent to download and install external software and configurations. This includes the open-computer-use package via npm and skill content from the iFurySt/open-codex-computer-use GitHub repository using npx.
[DYNAMIC_EXECUTION]: The tool features a JavaScript execution environment (ocu js and ocu repl) for automating complex sequences. This allows for the execution of arbitrary JavaScript code strings within the agent's operational context.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the host's desktop environment, including accessibility trees, application state snapshots, and text content (e.g., chat histories, emails). This data provides an attack surface where embedded instructions could influence agent behavior.
Ingestion points: Data enters the context through get_app_state, list_apps, and automated screen snapshots captured from the user's session.
Boundary markers: The instructions do not specify the use of clear delimiters or guardrails to help the agent distinguish between UI data and its own system instructions.
Capability inventory: The skill possesses significant capabilities, including file system modification (via installers), software installation, GUI manipulation (clicks, keystrokes), and arbitrary code execution.
Sanitization: No explicit sanitization, filtering, or validation of the desktop-derived content is mentioned in the provided documentation.
[PERSISTENCE]: The installation process for various agent platforms (such as Claude, Gemini, and DeepSeek Harness) involves modifying configuration files and writing to local directories (e.g., ~/.dsh/ocu-hooks.json and profile patch files) to ensure the tool persists across agent restarts and new sessions.

open-computer-use