How DOT404 works

The pipeline, step by step

1. You ask. Type in the main window, dictate, hold the mic on Amethyst, or ask about your screen. Nothing runs before you act.
2. Context is added — only what you opted into. A plain question travels as text. A screen question captures one fresh screenshot of the display your cursor is on. An attached file is parsed on your device and its extracted text rides along.
3. The request goes through one secure gateway. Signed in, your message streams to DOT404’s Supabase Edge Function, which verifies your session and forwards to the AI provider for this request type. Keys live server-side; the app never holds them. Security & trust
4. The answer streams back. Tokens render live in the conversation — Amethyst’s thread and the main window show the same exchange, because it is one conversation. You can stop generation mid-stream.
5. Guidance is drawn — optionally. If the question asks to point or click, a grounding pass localizes up to eight boxes on the same screenshot. A transparent, click-through overlay draws arrows, rectangles, circles, pointers, lines or squares on your display for eight seconds, then removes itself.
6. Lessons stay until you finish them. “Explain how to…” turns the answer into ordered steps with highlights that persist for the lesson instead of expiring after eight seconds.
7. Everything lands locally. The conversation is written to a SQLite database on your device. The screenshot is dropped after the answer — no files, no history, no thumbnails.

The parts that never happen

• No background capture. No periodic screenshots, no screen recording — there is no code path for it.
• No mouse or keyboard control. DOT404 points; you click.
• No always-on microphone. Recording starts at an explicit press and stops within 60 seconds.
• No chat content in the cloud. AI requests leave the device; your history does not get stored anywhere else.

When you are signed out

The app runs fully offline for character animation and local conversations. AI replies fall back to a local preview clearly labeled as mock — no network, no cost. Voice, screen and vision answers are signed-in features; they say so honestly instead of faking a result.