Skip to content
ScribeAId
IT

Open source · AGPL-3.0

The open-source clinical scribe.

ScribeAId records the visit, transcribes it while it happens and drafts a structured medical report. The clinician reviews, edits and signs. Run it on your own hardware and keep the audio where it belongs.

View on GitHubQuickstart

Built to be taken apart

Modular

Web app, desktop shell, backend and speech-to-text are separate components behind stable interfaces. Run all of it, or just the transcription service as an OpenAI-compatible endpoint.

On-premise

Clients, backend and speech recognition run on your servers. With self-hosted transcription the consultation audio never leaves your infrastructure.

Near real-time

Audio streams to the backend in chunks and comes back as text during the visit. The transcript is ready when the consultation ends, not minutes later.

Open throughout

Code under AGPL-3.0 and the speech model published on Hugging Face. No trial tier, no per-seat licence, no vendor holding your clinical data.

How it works

From the conversation to a signed report, with the clinician in control at the end.

  1. 01

    Record

    The web or desktop app captures the visit and streams audio over a websocket.

  2. 02

    Transcribe

    MedWhisper on your hardware, or any OpenAI-compatible speech-to-text endpoint.

  3. 03

    Draft

    An LLM of your choice turns the transcript into a structured medical report.

  4. 04

    Review & sign

    The clinician reads, edits and signs. Nothing is final until they say so.

Speech recognition

MedWhisper, released in the open

An Italian speech recognition model fine-tuned on clinical language and published on Hugging Face under the same licence as the code — not hidden behind an API.

Model on Hugging FaceCTranslate2 build
4.5%
word error rate on a held-out clinical test set, against 7.9% for the base model
−43%
relative reduction in errors over Whisper large-v3 turbo
0.5–0.7×
of the audio duration to transcribe on CPU

Data sovereignty

You decide where each piece of data goes

Consultation audio and clinical narrative are health data. ScribeAId gives you two independent knobs, because the two halves have very different profiles.

Transcription — the audio

Self-host MedWhisper and recordings never leave your network, or point it at an EU-resident endpoint. One environment variable, no code changes.

TRANSCRIPTION_BASE_URL

Report generation — the text

Send the transcript to the OpenAI-compatible provider you trust: Azure OpenAI in EU regions, Claude via Amazon Bedrock or Vertex AI, or your own gateway.

OPENAI_BASE_URL

Quickstart

Running in two minutes

You need Node.js, pnpm, uv and an API key for an OpenAI-compatible provider. The backend starts on :8099 and the web app on :3001.

Read the documentation
$ git clone https://github.com/ReportAId/scribeaid-oss.git
$ cd scribeaid-oss
$ cp .env.example .env  # then add your OPENAI_API_KEY
$ make dev
  • make dev-localadds MedWhisper on your machine — the audio stays local
  • docker compose up --build -dthe full stack behind one origin on :3000

Licensing

Two ways to use ScribeAId

Same code, same model. Pick the licence that fits how you run it.

Open source · Free

AGPL-3.0

For hospitals, clinics and developers who run ScribeAId themselves.

  • Full source code and the MedWhisper model
  • Use, modify and self-host it inside your organisation
  • Offer a modified version as a service to others, and you share its source
  • Community support through GitHub issues and discussions
Get the code

For vendors and integrators

Commercial licence

When AGPL does not fit — integrating ScribeAId into a proprietary product, for example — with direct support from the team that builds it.

  • No copyleft obligations on your product
  • Architectural support for your deployment and integration
  • Bug fixes and security patches
  • Review and support on your pull requests
Contact us