An anime companion that actually runs on your own PC

What it does

Amika is a Live2D anime character who talks with you by voice and text in real time, remembers past conversations, and knows your name and city. The whole stack — speech recognition, memory, text-to-speech, even camera-based “eyes” — runs on your own Windows machine. Only the language model is pluggable: either OpenRouter on your own key, or fully local via Ollama.

The interesting bit

The split is unusually honest about what local AI costs. Ears, voice, memory and eyes always run locally (~6 GB VRAM, ~10 GB with vision), but the README is blunt that a local brain needs a 27B model minimum — 70B for the “full, reliable experience” — and that anything smaller makes her lose character and break her reply format. It even publishes the math: ~$1.50/month at 30 conversation turns a day via OpenRouter, versus roughly 19–50 GB of VRAM to go fully local.

Key highlights

• Live2D avatar with real-time voice conversation and long-term memory of you
• Semi-local (OpenRouter) or fully local (Ollama) — no vendor lock-in for the brain
• Helper agents (Hermes Agent, nanoMuse, any MCP server) can run errands like web search and file work
• Optional extras: Discord/Telegram messaging, Tailscale access from your phone, Zigbee smart-home plugs (with a safety disclaimer)
• Resumable setup, per-component health checks, and a genuinely detailed troubleshooting table

Caveats

• Windows-only in practice; Linux is explicitly untested
• Heavy requirements: NVIDIA RTX 20-series+, 8 GB VRAM (16 GB for her eyes), ~45 GB disk
• Her prompts are written in Chinese, so the model must support it — some 70B models like llama3.3:70b won’t work
• Voice and Live2D assets are under a personal-use license, not fully open

Verdict

Worth a look if you’re curious how far a self-hosted AI companion can go, or you want a template for wiring voice, memory, and MCP tools together. Skip it if you’re on Linux, lack an NVIDIA card, or the anime-companion premise does nothing for you.