Skip to content

hearth for Windows

A desktop application that runs language models on your own GPU and an autonomous coding agent against your own files. No terminal, no account, no per-token bill, and nothing leaving the machine unless you approve it.

Start here

What it is

It brings its own engine

The installer carries a pinned CPython and llama.cpp’s llama-server. Nothing to install first, and no Ollama required. It borrows the WebView2 runtime Windows already ships rather than bundling a second browser.

It asks before it acts

Four permission modes. The default reads freely and stops before every file write and every shell command. Underneath, each run carries a fixed list of tools it may touch at all, enforced even in bypass mode.

It knows what fits

The model shop measures your RAM and VRAM and computes whether a given quantisation actually fits, KV cache included, rather than recommending something that will swap to disk and feel broken.

It stays inside the folder

Every file tool passes a containment check: path traversal, absolute paths, reserved device names, alternate data streams and NTFS junctions are all rejected, so a confused agent cannot reach outside the directory you named.

How it is built

Built by Eric Catalano. See About for the project, its licence, and how to report a problem.