The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs users to download and execute binary installers (OpenClawCompanion-Setup-x64.exe and OpenClawCompanion-Setup-arm64.exe) from the openclaw/openclaw GitHub repository.
[COMMAND_EXECUTION]: The "Node Mode" functionality implements a system.run command that allows the AI agent to execute arbitrary PowerShell commands on the Windows host.
[DATA_EXFILTRATION]: The skill provides the agent with capabilities to capture sensitive data from the host, including screen snapshots (screen.snapshot), camera images (camera.snap), audio transcription (stt.transcribe), and geographic location (location.get).
[DYNAMIC_EXECUTION]: The skill supports a canvas.eval command which allows the agent to execute arbitrary code within the companion application's WebView2 context.
[INDIRECT_PROMPT_INJECTION]: The skill creates an attack surface where instructions processed by the AI agent (e.g., from web content or chat) could be translated into local system commands via the WebSocket gateway.
Ingestion points: WebSocket messages received from the remote OpenClaw gateway as described in SKILL.md.
Boundary markers: The architecture requires device approval and relies on a local allowCommands configuration file to restrict agent actions.
Capability inventory: The skill grants access to shell execution (system.run), UI interaction (canvas.navigate), and various hardware sensors.
Sanitization: Command safety is dependent on the agent's logic and the user's manual configuration of allowed commands in openclaw.json.
[PERSISTENCE]: The configuration schema in SKILL.md includes an AutoStart option that, when enabled, ensures the companion application and its remote control capabilities persist across system reboots.

openclaw-windows-companion