Why avatar training and inference never leave your machine
The design thesis behind keeping voice and video avatar generation entirely local instead of uploading biometric data to a cloud service.
By Agent Software
Biometric data is not like other data
If a password leaks, you rotate it. If your face and voice leak, you can't. That distinction is why voice cloning and avatar generation deserve a different default posture than most SaaS features — and why "upload a sample, we'll process it in the cloud" is a much bigger ask for this category of product than it sounds.
The local-only pipeline
Agent Avatar's voice cloning and video generation pipeline runs entirely on local hardware. Fine-tuning a voice model happens on your own GPU. Generating lip-synced or animated video happens on your own machine. There's no step in the pipeline where a face or voice sample gets sent to a remote server for processing.
What that costs you
Being upfront about the tradeoff: local training takes real time and real GPU resources that a cloud service could otherwise absorb on your behalf, and it sets a real hardware floor — this isn't something that runs well on a laptop with no dedicated GPU. Local-first is a choice with a genuine cost, not a strictly-better option in every dimension.
What it buys you back
In exchange, the only place a trained avatar model exists is a machine you control. There's no server-side copy to be breached, no third-party retention policy to trust, and no scenario where a service shutdown or acquisition puts your likeness data somewhere you didn't choose. For data that can't be reissued if it leaks, that tradeoff is deliberate, not incidental.
How this compares to the cloud-avatar norm
Most competing services in this space are built around uploading a reference sample once and running inference in the cloud afterward — convenient, but it means your biometric data lives on someone else's infrastructure indefinitely. Agent Avatar's local-only design is a direct response to that norm, not an accident of how the pipeline happened to get built.
An honest status note
Agent Avatar is an internal project in development, currently used to produce the suite's own presenter and voice material rather than offered as a public, self-serve product. The local-only architecture described here is real and already in use internally — it's the availability, not the design, that's still ahead.
