How to Use Local AI During a Proctored Exam on Mac
Local AI means: no cloud, no API call, no traffic to OpenAI or Anthropic. The model runs on your Mac via Ollama, llama.cpp, or LM Studio. The Fracture overlay carries the chat UI for whichever you choose. The combination - invisible-to-capture overlay plus on-device LLM - is the most private setup possible.
Three paths to local: Ollama (one-click install, simplest), LM Studio (Mac app with model browser), llama.cpp (raw, command-line, fastest). All three expose an HTTP API on localhost that the Fracture overlay can point at. Mistral 7B and Llama 3.1 8B are the standard picks for routine homework; DeepSeek Coder for code. Pre-warm the model by running one prompt before the exam (cold-start can be slow). During the exam, the overlay carries the chat interface; the model runs on your Mac's Apple Silicon GPU. Nothing leaves the device.
7B local ~ GPT-3.5; cloud frontier ~ GPT-4. For routine homework explanation, local is fine. For hard problems, cloud is stronger.
It can enumerate processes and find Ollama by name. Use llama.cpp with a renamed binary if you want zero process-list footprint.
Yes if you have RAM. 16 GB Mac comfortably runs Mistral 7B + DeepSeek Coder 6.7B simultaneously.
Need reference access during your exam?
Fracture runs a native floating overlay that sets the macOS window privacy flag. It is 100% excluded from screen capture feeds, window lists, and proctor recordings.