Voice-to-voice AI assistant using Gemini Live API. Speak to the AI and get spoken responses. Use when you want to have natural voice conversations with an AI...

What to consider before installing: - Metadata mismatch: the registry metadata claims no required env vars, but skill.json and handler.py require GEMINI_API_KEY. Verify the source and ask the publisher to correct metadata before trusting the package. - Secrets: the skill will read a .env file in its directory and import values into the process environment if present. Do not put unrelated secrets in that .env file; only store the Gemini API key there if you accept the risk. - Network and privacy: the skill uses google-genai to connect to Google's Gemini service — any voice/text you send will go to Google's servers. If you have privacy concerns, do not use it with sensitive data. - Local files: the skill writes audio to /tmp/gemini_voice_<id>.ogg and removes an intermediate WAV file. OGG files may persist until cleared; consider automatic cleanup or a different output directory if multiple users share the system. - Dependencies and binaries: you must install the listed Python packages and ensure FFmpeg is available at the expected path (handler.py uses /usr/bin/ffmpeg). Confirm the google-genai package you install is the official one and review its network behavior. - Source trust: the skill has no homepage and an unknown source/owner. If you need strong assurance, request a verified source, a repository link, or an upstream release to inspect before running it with your API key. If those concerns are acceptable and you trust the publisher, the code itself is consistent with its stated functionality; otherwise treat this as untrusted until the metadata and provenance issues are resolved.

Type: OpenClaw Skill Name: gemini-voice-assistant Version: 1.0.0 The skill is classified as suspicious due to a significant prompt injection vulnerability in `scripts/handler.py`. The `system_instruction` parameter, which can be provided via `request_data` or CLI arguments, is directly passed to the Gemini API without sanitization or validation. This allows an attacker or a compromised OpenClaw agent to inject arbitrary instructions into the LLM's system prompt, potentially manipulating the AI's behavior, extracting sensitive information, or causing unintended actions. While the skill itself does not contain malicious instructions, this design flaw creates a critical attack surface.