saferowThree jobs
saferow names the jobs the same way everywhere (Models, ⇧⌘M, and Settings › AI):
- Model answers every question: it plans, runs the queries, proposes changes and writes the answer.
- Helper does small, fast jobs: conversation titles, the echo line under your question, sorting a question, filter suggestions. Left unpicked, the Model does them.
- Table finder picks the tables to send when a schema is too big to send whole (over 9,000 characters).
Each can be Automatic, which picks the best model this Mac has, by saferow’s own tests. Automatic never picks a cloud model.
Where models run
- On this Mac, by saferow. saferow runs GGUF models itself with Metal; no other app is needed, and a model unloads after 5 idle minutes. Models marked “verified here” passed saferow’s self-check and agent loop.
- On this Mac, by another app. Ollama, LM Studio and oMLX are found and used. A model you already have in Ollama or LM Studio is used without a second download.
- On your network. Any OpenAI- or Anthropic-compatible server: Settings › AI › Other servers.
- In the cloud. OpenRouter with your own account (signing in opens your browser), or your own key from Anthropic, OpenAI, Google Gemini, xAI or DeepSeek. Keys go straight to your Keychain.
What saferow suggests
| Mac memory | Model | Helper |
|---|---|---|
| 32 GB or more | Qwen 3.6 35B-A3B | Granite 4 7B-A1B |
| 16 GB | Qwen 3.5 9B | Granite 4.1 3B |
| 8 GB | Qwen 3.5 4B | Granite 4.1 3B |
Get models (in Models) searches Hugging Face only when you type, and shows a fit estimate for your Mac before anything downloads. The context ring under the composer shows how full the model’s window is; near the top, start a new conversation with a summary.
What goes to a cloud model
The question, the parts of the schema it needs, and the rows it reads, under your own account or key. Values in columns saferow recognises as personal data are withheld from any model off the Mac; before such a question leaves, saferow offers a model on this Mac, the schema only, or masked columns. Set Keep rows on this Mac on a connection and the kernel refuses any model off the Mac for it. Settings › Privacy › Only models on this Mac switches every cloud provider off at once.
The line under each answer says who answered, where, and how much left the Mac, such as “1.2 KB left this Mac · schema only”.