The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill automates the installation of the agent-click utility from the public npm registry to the host machine if it is not already present.
[COMMAND_EXECUTION]: Executes various shell commands through a local tunnel (agent-tunnel) to facilitate desktop interactions, including opening applications, managing windows, and simulating keyboard/mouse events.
[REMOTE_CODE_EXECUTION]: The skill enables the agent to execute complex UI automation sequences and shell-based CLI commands on the user's computer to fulfill desktop automation tasks.
[PROMPT_INJECTION]: The skill is designed to read and process content directly from the user's screen (via snapshots and value retrieval), which creates a surface for indirect prompt injection if the agent encounters malicious instructions embedded in web pages, documents, or other UI elements.
Ingestion points: Screen content and UI element attributes captured via the snapshot, text, and get-value commands in SKILL.md.
Boundary markers: The instructions do not specify delimiters or constraints for the agent to distinguish between valid UI data and potential instructions found within that data.
Capability inventory: Extensive desktop control including shell command execution, file manipulation via the GUI, and application interaction through the agent-tunnel primitive.
Sanitization: No sanitization or filtering of screen-captured text is performed before it is added to the agent's context.

computer-use