kevMac

kevMac is a lightweight macOS app for the kev family of decision models. Paste any document — a customer message, a review, an article — build questions as simple forms, and get calibrated answers with animated probability bars. The app converts your form into the engine's JSON contract automatically, and turns the JSON result back into plain language. No terminal, no JSON, no button presses on a website. One setup button, and everything runs locally on your Apple Silicon Mac.

Features:

• Plain text in — Type or paste a document and build questions as simple forms. Each question picks a type: Yes / No (with optional "what Yes / No means" descriptions), Multiple Choice (named options with optional descriptions), or Rating (ordered levels, worst → best).
• JSON, invisibly — The form is converted into the kev /v1/systemone contract automatically. You never see or write JSON.
• Normal text out — Answers render as plain language — "Returns — 83% confident" — with animated probability bars (critically damped springs), confidence captions, and latency + token usage. A Copy Report button puts a plain-text summary on the clipboard.
• Add / remove everything — Questions, choice options, and rating levels each have add and remove buttons, validated inline (a choice needs at least two named options, a rating at least two levels).
• Model selection — The Model menu in the toolbar lists every supported checkpoint. The setup installs kev-0.8b by default; already-downloaded models say so, and larger ones (kev-4b, kev-9b) show their download size — pick one to download it on demand, and the engine restarts on it automatically.
• Examples — Three presets: Support triage (the shoes-arrived-late example), News article, and Review rating.
• One-button setup — The first launch installs everything into a dedicated ~/.kevMac folder automatically: uv, Python 3.13, the kev engine, and the model weights. The weights download by themselves — no button press.
• Runs fully offline after setup — The engine serves from ~/.kevMac/Cache, starts on launch, health-checks, and warms up in the background so your first real question answers fast.