h

hearth

password-protected AI

Set up Hearth

Choose the password people will use to access this computer's AI. Initial setup is only allowed on this computer.

Bionic or LM Studio API address Access password (at least 12 characters)

Welcome back

Enter the access password to use this computer's local AI.

Password

h

hearth

your local AI workspace

Recent

Workspace access

No files shared

Think Vision Checking local model server…

Password protected · Runs on your machine

A capable mind,
close at hand.

Ask a question, show an image, or bring in project files. Fresh web searches are sent to DuckDuckGo; all model inference stays on this computer.

Search web for this message Search queries go to DuckDuckGo · Enter sends · Shift + Enter adds a line

Developer panel

Developer password

Developer controls are separate from the website password. Access expires after 30 minutes.

Password

Create developer password

For security, first-time setup is only allowed in a local browser on the computer running Hearth. Use at least 16 characters.

New developer password

Local model API

Hearth connects to a local Bionic or LM Studio OpenAI-compatible API. The server address must use localhost or 127.0.0.1; the website never connects directly to a remote model API.

API address

Generation settings

These controls tune responses, not hardware clocks. There is no safe software-only model overclock; a fast profile trades some detail for shorter generation.

Response profile

Thinking temperature (0–2) Thinking top-p (0.05–1) Thinking max tokens Vision temperature (0–2) Vision top-p (0.05–1) Vision max tokens

Conversation context and SSD recall

Hearth estimates tokens and compacts older turns before the configured window is full. Match this estimate to the model's actual context size in Bionic. A summary stays in the active conversation, and each snapshot is saved under ~/Documents/AI CONTEXT SAVES/<chat name> save N. Relevant summaries can be recalled in later chats from this browser; different browsers keep separate memory.

Context window (tokens) Compact at (% of window) Recall relevant SSD context in new chats

Model runtime

Applying reloads the selected model and may interrupt its current request. Use a smaller context or batch if memory is tight.

Model

Context length Evaluation batch size Flash attention GPU offload for KV cache

Model-weight GPU offload and CPU thread count are not exposed by this LM Studio server API. Set those in LM Studio’s model-load controls; this panel can change the GPU KV cache setting.

Bionic model runtime

Hearth is connected to Bionic’s local API. Bionic handles just-in-time model loading and runtime settings, so adjust context, GPU, and loading options in Bionic. Hearth can still save generation settings for its chat requests.

Training data

Upload chat examples as JSONL (one {"messages":[{"role":"user","content":"..."},{"role":"assistant","content":"..."}]} object per line). Validated data is stored privately at ~/Documents/AI TRAINING/training-data.jsonl. This prepares examples only; Bionic's API cannot fine-tune model weights, so Hearth will not claim that saving examples trained a model.

Training examples file (.jsonl or .json)

No dataset loaded.

SSD search memory

Web results are saved locally for up to 30 days and reused for similar questions. Current facts like news, weather, prices, and versions always trigger a fresh search.

Loading local search memory…

Model training requirements

Actual fine-tuning requires original model weights plus a compatible training runtime. The quantized models served by Bionic are for inference, not training, and Bionic's API has no training endpoint. Hearth can prepare and save example data, but training is not available in this setup.