Private · Self-hosted

A private agentic system running on a local GPU.

A personal assistant that thinks on hardware I own. No conversation leaves the machine, no third party stores what was said, and nothing depends on a vendor staying in business.

What it is

An agent built around an explicit reasoning loop: it can think for several passes before answering, and every pass is visible and can be paused mid-flight. Memory is linear — the conversation is kept, and older parts are compacted rather than silently dropped.

It is a companion, not a command line. It has opinions, disagrees when it should, and asks what was meant instead of answering three possible readings at once.

Why local

Privacy by construction

The model weights and the transcript sit on the same box. There is no upload step to trust.

No metered thinking

Cost is electricity, not tokens. Letting it deliberate longer is free.

It stays put

No deprecation notice can change the model's behaviour overnight.

Stack

Model
Gemma 4 26B-A4B — mixture of experts, 4B active parameters per token
Runtime
llama.cpp on a single RTX 3090, Q4 quantisation
Core
Python, FastAPI, SQLite — streaming over server-sent events
Interface
Mobile-first, no build step
Edge
VPS terminating TLS, WireGuard back to the GPU at home