The Agent Skills Directory

Data Transmission to External Service: The skill transmits screen descriptions, including element names, roles, and up to 60 characters of field values, to an external API (defaulting to api.typesafe.ai). While this is necessary for the AI to make decisions, it may include sensitive information from open documents or forms.
Indirect Prompt Injection Surface: The skill automates desktop actions by reading screen content, which constitutes untrusted data. If a window or document contains text designed to influence an LLM, it could potentially trigger unintended actions such as clicking or typing.
Ingestion points: Screen descriptions are read and processed by scripts/jev/screen.mjs and sent to the decision endpoint.
Boundary markers: The documentation does not specify the use of strict boundary markers to distinguish screen content from agent instructions in the prompt.
Capability inventory: The skill can perform diverse desktop operations including CLICK, TYPE_TEXT, CHECK, SCROLL, and set-value via the agent-desktop binary.
Sanitization: The skill redacts explicitly supplied text from some logs, but live screen content remains a vector for instructions.
External Dependency Requirement: The skill requires a local clone of the agent-desktop repository and its associated binary to interface with the operating system, which is a common pattern for desktop automation tools.

jev-desktop