🛡️ Sovereign & Local-First AI

The First Open-Source
Self-Hosted AI Companion

Engage with Hasaji, your customized local companion. Haven is the first open-source, fully self-hosted AI companion system, bringing together interactive desktop overlays, local image generation, and offline Piper speech on your own hardware.

Hasaji Companion Portrait

The Haven Architecture Suite

A synchronized network of local applications designed to host, render, and orchestrate secure AI companion systems.

🖥️ Desktop client (Testing Phase)

Interactive Companion

The companion UI floats natively on your desktop. Currently in active development and testing, featuring automated screen capturing, monitor-aware bounds mapping, and smooth status-pulsing borders.

⚙️ C# orchestrator

Haven Server Core

A hardened C# .NET 10 orchestrator. Handles OAuth identity mapping, JWT-secured WebSocket connections, rate-limiting, and prompt-injection defense mechanisms while logging actions in a central database.

🎨 local sd-server

SD Portrait Engine

Generates custom avatars dynamically. Integrates FaceDetectorYN and GFPGAN face restoration pipelines in ONNX to automatically clean, align, and reconstruct lifelike eyes and facial details.

Gemma 4 Turbo

Runs the #1 community-compiled Google Gemma 4 model on the Ollama registry. Built from native 16-bit float source weights using IQ4_XS non-linear importance matrix quantization.

Offers custom mixtures of merges with Mahou roleplay weights, preserving full vision projector layers for multimodal understanding while delivering high-speed execution directly on commodity CPUs.

>35K
Downloads
10.4t/s
Xeon CPU Speed
Ollama Engine Console
ollama run ssfdre38/gemma4-turbo
pulling manifest ... done
loading model weights ... 100%
merging vision projector tensors ... 100%
>>> CPU execution loaded (8 threads, 10.4 tok/s)