zeekk0/OpenAmika

What it is

OpenAmika packages an AI companion character, Amika, as a Windows application. She is a Live2D avatar who talks via speech and text in real time, keeps long-term memory of the user, and can be configured to live in a chosen city. The language model can run through OpenRouter using your own key or fully locally through Ollama; speech recognition, voice synthesis, memory, and vision run on the local machine. The project targets people who want a private, self-hosted companion rather than a browser-only service.

The repository contains the Python application, setup scripts, a web UI served at http://localhost:53000, and optional integrations for helper agents and smart-home devices. A hosted version also exists at mekahime.com/amika, and the project site is mekahime.com/openamika.

Quick start

Requirements: Windows 10 or 11, an NVIDIA RTX 20-series or newer GPU with 8 GB VRAM (16 GB if you enable her camera eyes), Python 3.11 on PATH, Git, and roughly 45 GB of free disk. Setup downloads about 15–20 GB and takes 30–60 minutes. For the language model you need either an OpenRouter API key or Ollama with a model of at least 27B parameters.

In a normal PowerShell window:

cd C:\
git clone https://github.com/zeekk0/OpenAmika.git
cd OpenAmika
powershell -ExecutionPolicy Bypass -File scripts\setup_windows.ps1
envs\brain\Scripts\python.exe scripts\setup_tts.py
.\Start_OpenAmika.bat

The voice and Live2D model are downloaded only after you accept MekaHime's personal-use license by typing yes when prompted. Each setup command ends with a summary; OK means a component is ready, and a second run resumes and skips finished parts. After startup your browser opens at http://localhost:53000; the first visit asks for your name, city, and OpenRouter key if none is configured. Click the page once so the browser allows audio, then press the power button and speak or type. To verify the installation without starting the UI, run envs\brain\Scripts\python.exe launch.py --check; it ends with result: ready or names what needs fixing. A troubleshooting table covering missing Python, GPU detection, download and voice issues is in the README.

Choosing a language model

OpenRouter is the recommended option: responses are faster and cost roughly $1.50 a month for about 30 turns a day, or about $5 for 100 turns. Fully local Ollama needs a model of 27B parameters or more, with 70B or more for the full experience; smaller models may lose character consistency, forget context, or break the reply format. A local model shares VRAM with the local speech, memory, and vision components, so check that it fits alongside them. To switch, install Ollama, pull a supported model such as gemma3:27b, then choose Brain → Ollama → Detect models → Use this model on the Setup page.

Constraints and licensing

Linux support is untested. Without an NVIDIA GPU, setup can run in CPU mode via -File scripts\setup_windows.ps1 --cpu, but performance will differ. The application's own code is MIT-licensed, while the character, art, voice, Live2D model, and related weights are not: they belong to MekaHime Inc. and are free only under a personal-use license (18+, no redistribution or re-hosting, no training other models on them). Limited social-media monetization and fan merch under US$1,000 are allowed under the stated creator rules; other commercial use requires a license from MekaHime. The Hugging Face license governs those assets.

Optional features have their own conditions. Smart-home control stays disabled until you accept the IoT safety statement, and the README warns never to connect heaters, locks, valves, or other harmful devices. Helper agents such as Hermes Agent and nanoMuse are separate programs installed by the user. Discord/Telegram messaging and Tailscale access require their respective configuration.

HotRepo's take

On the HotRepo new-project list recorded on 2026-11-05, OpenAmika gained 23 stars and 4 forks in a one-hour window, reaching 132 stars and 17 forks by the end of that observation. That single short snapshot shows early attention but is too brief to describe a trend. The project's distinguishing choice is its explicit split between local speech/vision/memory and a configurable language-model backend, plus unusually detailed hardware and licensing documentation for a companion app.