The Agent Skills Directory

[COMMAND_EXECUTION]: The skill documents and provides code examples for executing arbitrary bash commands and other binaries on remote cloud computers via dedicated API endpoints (/bash, /exec). While intended for remote management, this provides a direct path for command execution on target infrastructure.
[EXTERNAL_DOWNLOADS]: The instructions involve downloading external software, including pre-built binaries (OS1.app.zip) and cloning source code from GitHub (nickvasilescu/hermes-desktop-os1). It also manages the automated installation of the "Hermes Agent" on remote virtual machines.
[INDIRECT_PROMPT_INJECTION]: The skill creates a surface for indirect prompt injection as it processes session history and file contents from remote computers.
Ingestion points: Session metadata, message content, and remote files accessed via the file editor (SKILL.md).
Boundary markers: None identified in the documentation for delimiting untrusted data from agent instructions.
Capability inventory: The skill has extensive capabilities including bash execution, arbitrary command execution, file writing, and network operations to third-party APIs (SKILL.md).
Sanitization: No specific sanitization or filtering of external data is mentioned, though file conflict detection is implemented.
[CREDENTIALS_UNSAFE]: The skill handles highly sensitive credentials, specifically ORGO_API_KEY and OPENAI_API_KEY. While it recommends using the macOS Keychain for secure storage, the documentation provides examples of these keys being passed through environment variables and raw HTTP Authorization headers.
[DYNAMIC_EXECUTION]: The voice mode configuration utilizes npx -y @orgo-ai/mcp, which downloads and executes the latest version of a tool from the npm registry at runtime.

hermes-desktop-os1-native-macos-client