From a press of a button to a searchable, summarised transcript, every step runs on hardware and services you control. Here is the whole workflow, end to end.
The device is a small ESP32-S3 board with a digital MEMS microphone and a microSD card. A single button starts and stops recording; a status light shows what it is doing. Audio is written straight to the card as it is captured, so there is no length limit and nothing needs a network connection just to start recording. Each recording is saved with a small metadata file (duration, sample rate, and a unique id) so the rest of the pipeline can track it.
Nightjar has three independent ways to move recordings off the device, so one of them always fits your situation:
Once a recording reaches your server, it is transcribed into accurate, timestamped text. The transcription engine is your choice: a fully local model for maximum privacy, or a commercial speech-to-text API for the highest accuracy. Nothing is hard-wired to a single vendor.
The transcript is then passed to a language model, which writes a structured summary: a short overview, the key points, and any action items. Again, the model is yours to pick, from a local model running on your own machine to a frontier commercial API. See Bring your own AI for the full list.
Everything is available through a completely open REST API. Pull the original audio, the plain-text transcript, subtitle files (SRT or VTT), a Markdown summary, or the full record as JSON. Because the API is documented and open, you can wire Nightjar into whatever you already use.
Four parts, all under your control, connected by an open API and a shared data model.
The guides walk through connecting an AI provider, self-hosting the server, and building the device yourself.