Ask us
All work

An AI assistant where no AI provider ever sees the chats.

What
Self-hosted AI assistant
For
One owner, at home
Built with
Python, TypeScript, llama.cpp, ComfyUI, passkeys
Status
Early: 3 of 8 phases done
Code on GitHub

Ask it to do something

This tool is set to

clockLook

The rules are Hearthkeep's own permissions table. clock and note are its real demo tools; outward tools arrive in Phase 4.

Hearthkeep on a phone: a sales file is attached, and the local model finds a 13-day gap in March, every number computed in the sandbox

The problem

Cloud assistants see everything you type, and an agent that can act needs rules it cannot talk its way around.

What it was

Local models on your own PC, passkeys, and nothing leaves without your yes.

How it was done

  1. Local models only: chat through llama.cpp, images through ComfyUI, swapped on the GPU when they don't fit together.
  2. Every tool call is Look, Change local or Outward, and every tool is Allowed, Ask or Off. Unknown tools fail closed.
  3. Anything outward waits for an approval and a fresh passkey check. The model cannot approve its own requests.
  4. Everything listens on 127.0.0.1; phones reach it only over a private Tailscale network. Sign-ins, approvals and permission changes go into a log the app cannot edit.

The result

Phases 1 to 3 done: passkeys, approvals, audit log, local chat, image generation and photo editing, and files analysed in a sandbox where every number is checked by code. Public on GitHub under Apache 2.0.

Want something like it?