The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructions require the execution of shell scripts (scripts/start.sh or scripts/start.bat) and various openclaw CLI commands to manage devices and agents. It also uses ffmpeg (via imageio-ffmpeg) for audio processing with configurable filter chains, and allows overriding the binary path via the FFMPEG_BIN environment variable.
[EXTERNAL_DOWNLOADS]: The installation workflow involves cloning an entire external code repository from https://github.com/RubinTry/assistant-x-openclaw.git. Additionally, it prompts the user to download several large binary machine learning models (ASR, KWS, TTS, and Speaker Verification) from external sources referenced in the repository's README.
[CREDENTIALS_UNSAFE]: The assistant explicitly requests the operator.admin scope during the OpenClaw device pairing process, which provides administrative control over the gateway rather than just message access. It also generates and stores a private keypair at ~/.openclaw/devices/voice_assistant_keypair.json for authentication.
[DATA_EXFILTRATION]: The skill establishes a persistent WebSocket connection to an external or local gateway (OPENCLAW_GATEWAY_URL) and transmits voice-derived text data. While intended for LLM processing, this represents a continuous data stream to a network endpoint.
[INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted audio input from the environment, converts it to text via local ASR, and pipes it into an LLM gateway. There are no documented boundary markers or sanitization steps to prevent malicious voice commands from influencing the assistant's behavior or accessing connected tools.

assistant-x-openclaw-jarvis-voice-assistant