Installation

Agent Avatar is not a one-click installer, and we would rather say so here than have you discover it at step four. Installing means creating three conda environments and downloading several gigabytes of model weights. Budget an hour, most of it unattended downloading.

The current release is a qualified technical beta. Access is reviewed by email, and setup is intended for Linux or WSL2 users with a supported NVIDIA GPU. This page describes the install as it works today, so you know what you are signing up for.

Why three environments

The three engines have genuinely incompatible dependencies, so each gets its own conda environment:

The voice-clone environment is the one you interact with; it drives the others.

Steps

1. Install Miniconda and make sure conda is on your PATH.
2. Check your GPU is visible. Inside WSL2, nvidia-smi must list your card. If it does not, fix that first — nothing below will work without it.
3. Run the setup script. It creates the three environments, installs each engine's dependencies, and clones the third-party engine repositories.
4. Download the models. Several gigabytes of weights: F5-TTS, MuseTalk, SadTalker, Whisper for transcription, and the face restoration models. This is the long step. It only happens once.
5. Verify. ./avatar doctor checks every environment, model file, and GPU binding, and reports what is missing. --strict makes it exit non-zero, which is useful in a script.

After installing

The model files and environments are stored on your disk, and the local engine can perform generation without uploading source media for processing. The application may still contact network services for entitlement, updates, support, or optional integrations.

Next: Quickstart.