An AI assistant that runs on your own machine
Short answer
Nightjar's assistant runs an open-weight model from a file on your disk. Page content, your questions and attached files are processed locally and are not sent to any server.
The problem it exists for
Every useful browser assistant on the market answers by sending the page to a company. That is fine for a recipe and unacceptable for a client file, and there is no setting that changes it.
How it works
- The model is a GGUF file stored on your machine, run through llama.cpp on your GPU or CPU.
- Eleven models ship in the catalogue, from 0.75 GB to 30.4 GB, and any other GGUF file works.
- There is no account, no API key and no request. Switch off the network entirely and the assistant still answers.
What it does not do
- A model small enough to run locally is weaker than a frontier model on hard reasoning and obscure facts.
- It will not silently escalate a hard question to a cloud model, so sometimes the answer is that it cannot.
- The first run downloads a model, which needs a connection once.
Other ways people ask this
- browser local ai assistant