Product
The product
One desktop app. A Rust engine wires capture, speech-to-text, turn detection and an LLM together through plugins; a React interface renders the result as a live record with an answer column beside it.
Interview mode
Interview mode runs during a live conversation. The engine captures You and Them as separate streams, transcribes both, and detects when Them finishes speaking (an End of utterance).
You fire an Action when you want help: Ask, Suggest, Screen AI or Notes. Answers stream token by token in the Teleprompter style: first person, short spoken lines, the answer first. A bounded Rolling history of recent questions and answers gives follow-up questions context.
Auto-fire decides what happens at Them’s End of utterance: Off, Shadow (shows a ready cue only) or Live (requests an answer). Speculative answers can start from Them’s interim text, and Coach hints are local observations about your pace, pauses and filler words that never quote the transcript.
Notetaker mode
Notetaker transcribes both sides into a stored session. Afterwards you can Enhance a session into a summary, optionally following a template, export it as Markdown or PDF, search all sessions, and chat over sessions with answers that name the session each claim came from.
Only one mode runs at a time.
Plugins
Every stage is a plugin crate that depends only on contracts, the crate of shared traits and types, so a plugin cannot import a sibling. Audio sources: a virtual audio device via cpal (for example BlackHole), ScreenCaptureKit, and a shared capture with echo cancellation. Speech-to-text: Deepgram streaming or Apple on-device speech recognition. Turn detection: an energy-based detector. LLMs: Anthropic Claude, OpenAI and xAI Grok.
The interface
The window is a fixed-region panel: a status header, the Them and You panes, and the Answer column, with a Settings drawer for devices, providers, guidance and connectors. It is set in one monospace face, like a stenographer’s monitor, and every state change also has a non-colour signal.
Global shortcuts toggle the overlay window (Cmd+Shift+Space) and fire Ask (Cmd+Shift+Enter).
Use it openly
dont-panic records the other party’s voice. Recording and transcription laws differ by place, and many require everyone on the call to agree. Tell the other people in the conversation and get their consent where it is required.
Many interviews, exams and assessments do not allow AI assistance. Follow the rules of the conversation you are in. Use it where assistance is allowed or disclosed, for practice, or for your own meetings.
The app has a Stealth window setting that asks macOS to keep the window out of screen sharing and recording. It is best-effort only: on macOS 15 and later, screen sharing that uses ScreenCaptureKit still shows the window. It is not a way to hide your use of the tool, and this site does not offer it as one.
At a glance
- Platform
- macOS on Apple Silicon only; no Windows, Linux or mobile
- Users
- One person on their own Mac; no accounts, no sign-in
- Language
- English only for now
- Audio contract
- 16 kHz mono 16-bit PCM in 20 ms frames
- Cloud services
- Only the providers you configure, with your own API keys (Deepgram, Anthropic, OpenAI, xAI)
- Distribution
- Not distributed; private repository, access by request
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.