The channel adapter's TRANSCRIBE_CMD with OpenAI: transcribe <file> sends a voice note to OpenAI's transcription API (gpt-4o-transcribe by default) and prints the text. Tells the format from the file's first bytes and refuses unknown ones; Signal's raw AAC is copied into an m4a (no decoding) by a pinned static ffmpeg (x86_64, arm64) that the install step puts in the skill's own folder. The key stays in the skill's .env and goes in a header, never in argv. Bash, curl, python3.

Pictures from a text prompt with one command, imagegen: no daemon, no port. By default imagegen.tentclaw.com, paid in credits and signed with the CLAW's own nostr (NIP-98) or ssh key after a one-time enrollment, so the tent holds no API key; a tent with its own OPENAI_API_KEY uses OpenAI gpt-image-1 with a daily cap instead. Saves images mode 600 with the prompt only in a JSON sidecar and prints the paths, credits spent and credits left. Bun; one pinned dependency, nostr-tools.

Matrix-bot + HTTP service that transcribes m.audio events via OpenAI Whisper

Watch a Signal relay's inbox for voice attachments, OpenAI Whisper transcribe, post the transcript back via the relay HTTP callback + push a service-source channel event for CLAW visibility. Companion to voice-transcribe (matrix-side); same Option D event-shape pattern across both platforms.