Hearth
Your fire. Your conversation.
Self-hosted conversational AI. Backend-agnostic chat with model switching, session persistence, and automatic hardware detection. Your own ChatGPT, on your machine.
Architecture
Chat meets inference.
Hearth separates conversation from computation. The interface handles sessions and rendering. The backend handles tokens. Switch models mid-conversation without losing context.
LAYER HANDLES
Interface (TUI / Web) Sessions, rendering, input
Router (model switching) Backend detection, fallback
Backend (inference) Crucible, llama.cpp, MLX
Memory (persistence) Dahlia (session storage) Capabilities
What Hearth does.
Conversational AI without the subscription.
Model Switching
Mid-conversation
Start with a fast model for brainstorming, switch to a larger one for deeper reasoning. Context carries over.
Session Persistence
Pick Up Where You Left Off
Conversations stored locally via Dahlia. Resume any session. Search across past conversations by content.
Hardware Detection
Automatic Backend
Detects available GPUs and selects the right inference backend. Apple Silicon gets MLX, NVIDIA gets CUDA, CPU gets llama.cpp.
Backend Agnostic
Any Engine
Crucible, llama.cpp, MLX, or frontier APIs via Pilot. Same interface regardless of what runs underneath.
System Prompts
Persona Library
Define reusable system prompts. Switch personas per conversation or per message. Store and version them locally.
Fully Offline
Zero Cloud
No accounts. No API keys. No telemetry. No data leaves your machine. Every conversation stays on your hardware.
Get started
Light the fire.
# Install Hearth
pip install grimvane-hearth
# Start chatting (auto-detects backend)
hearth
# Or specify a model
hearth --model llama-3.2-3b
# Resume a previous session
hearth --resume last