This is a durable gateway demo — every leg needs a real model: real transcription, real translation, real voice synthesis. A sandboxed client can’t fake any of them. The artifact (the re-voiced clip) is re-hosted on our store, so the URL is durable.
The prompt
What runs
POST /fal/fal-ai/whisper— transcribe the source clip to text (~$0.02)POST /deepl/v1/translate— translate the transcript, context-aware (~$0.02)POST /elevenlabs/v1/text-to-speech/:voiceId— synthesize the translated voice (~$0.10)
Run it
Claude Desktop (MCP)
SDK
Expected output
Extend it
- Target any of DeepL’s 30+ languages by changing
target_lang(FR,DE,JA, …) - Swap DeepL for Google Translate (
/translate/v1/translate) for 130+ languages - Generate captions too: feed the transcript to Claude (
/anthropic/v1/messages) for an SRT - Loop over several
target_langvalues in parallel to ship one clip in five languages
