How chat sharing works

Sharing is off until you turn it on. When it is on, your chats with the agent help us find where local models and the harness fail, and, only if you allow it separately, train models. Here is exactly what happens.

Turning it on

The app asks once, on the first start, and the choice is in Settings, Data sharing, at any time. There are three boxes:

• Share my chats to improve Local Quench. Chats from this moment on.
• Also chats from before I turned it on. Off unless you tick it.
• Also allow training models with my shared chats. Off unless you tick it.

Each change is written to a log with the time and the version of the terms, on your computer and, with the next upload, on our server.

What is sent

For each chat: your messages, the model's replies and reasoning, the tool calls and their output, the model name, the harness settings, timings, and which parts the harness itself added (so training can tell your words from ours). Not sent: your files themselves (only what appeared in the chat), your settings, API keys entered in Settings, benchmark chats.

What is removed first

Before a chat is queued, the app removes, on your computer:

• API keys and tokens of known services (OpenAI, Anthropic, GitHub, Hugging Face, AWS, Google, Slack, Stripe and others), private keys and JWTs;
• values assigned to names like password, secret, token or api_key, and passwords inside URLs;
• long random strings that look like secrets;
• e-mail addresses;
• your home folder in paths (/home/jan becomes ~).

Settings, Data sharing, "Preview on my chats" shows any of your chats as the original next to what would be sent, with each removal and its reason. It cannot catch everything, such as names in code comments or customer data in a database dump. If a project has data like that, keep sharing off while you work on it.

When it is sent

Every five minutes the app looks for chats that changed and have been quiet for five minutes, puts them in a local outbox and uploads what is there. Without a connection, nothing is lost: the outbox waits and goes up the next time. A chat that continues later is sent again in full and replaces the earlier copy.

Who you are to us

A random install id, created by the app, and a secret that proves later uploads come from the same install. No account, no name, no e-mail. Your id is in Settings, Data sharing: give it to us if you write about your data, since we have no other way to find it.

Stopping and deleting

Untick the first box and nothing more is sent from that moment; the outbox is not uploaded. "Delete my shared data" empties the local outbox and sends a deletion request: the server removes everything stored under your install id, normally within minutes of the next connection.

Two separate choices

• Share my chats to improve Local Quench. We use them to find where models and the harness fail, to test changes and to build new benchmark tasks.
• Also allow training models with my shared chats. Optional. Only with this tick may your chats be used to train or fine-tune models, which may later be published as downloadable weights.

Each shared chat carries the training choice you had when it was sent, and your consent log records every change. Taking the training permission back applies to all your shared chats, earlier ones included, for any training that starts afterwards. A model trained before that keeps what it learned, because a trained model cannot forget single examples.

Benchmark results

A separate choice, for the public leaderboard. After every finished QuenchBench run the app asks whether to share that result, and shows everything that would go: model, quantization, processor, graphics card, memory, benchmark settings and per-task numbers. Never code, chats, file paths or commands. The tick in Settings, Data sharing only decides which answer is pre-selected in that question. Results from many computers are combined per model, quantization and hardware, so the leaderboard shows what really runs on a given card.

The terms you agree to

The app shows the full sharing terms before the box can be ticked: you allow us to store, analyse and modify the chats you share to develop Local Quench, free of charge, non-exclusively and worldwide, and to train models with them only if you gave the second yes. The chats themselves are never published, sold or given to other companies for their own use. Share only chats you are allowed to share: work for an employer or a client may need their permission. If the terms change, the app asks again and sends nothing more until you agree.