How it works
How it works
The engine runs one pipeline per side. Every stage is a plugin, and the canonical audio contract between them is 16 kHz mono 16-bit PCM in 20 ms frames.
- Step 1
Capture two sources
Your microphone becomes You. System audio, from a virtual device such as BlackHole or from ScreenCaptureKit, becomes Them. A side with no audio is filled with silence so the timeline stays continuous.
- Step 2
Transcribe each side
Each side streams to the speech-to-text plugin you chose: Deepgram (a cloud service, needs your key) or Apple on-device speech recognition (no key, no server). Interim text arrives first; final text replaces it.
- Step 3
Detect the end of a turn
An energy-based turn detector watches Them and marks each End of utterance. Auto-fire uses it: Off, Shadow shows a ready cue, Live requests an answer.
- Step 4
Fire an Action
Ask, Suggest, Screen AI or Notes, from the interface or
Cmd+Shift+Enter. The request carries the recent transcript, the Rolling history and your Knowledge base as guidance notes. - Step 5
Stream the answer
The LLM you chose (Anthropic Claude, OpenAI or xAI Grok) streams the answer token by token into the Answer column, in short first-person lines.
- Step 6
Keep what you choose
In Notetaker mode the session is stored in a local SQLite database you can summarise, export, search and chat over. In Interview mode you save a transcript only when you choose to.
What runs where
On your Mac: audio capture, Apple on-device speech recognition, turn detection, Coach hints, the session store and the search index.
At the providers you configure: Deepgram receives audio for transcription; Anthropic, OpenAI or xAI receive the transcript context and guidance for an answer, and a screenshot only when you fire Screen AI. MCP connectors you add receive the requests the app sends them.
Nothing is sent to the authors of dont-panic. There is no telemetry server.
A personal tool, shared by request
dont-panic is built for one person on one Mac. There is no download page and no account. The source repository is private; ask for access.