Deployment postures
at the edge.
One application, three ways to run it. The product doesn’t change; the network around it does, and so does exactly what we can claim.
Posture 01 Single device
One laptop, no network.
The whole pipeline runs on one managed laptop.
- Install the app and download models while connected
- Disconnect: local mode needs no network after setup
- Switch usage analytics off in Settings
- Block the update check at your firewall or through device management
What holdsIn local mode, meeting content never leaves the device.
Posture 02 Isolated network
Your model server, inside your boundary.
Endpoints record and transcribe locally. Transcript text, never audio, goes to a model server you run on the same network, such as Ollama or an OpenAI-compatible endpoint, so larger models can run on your GPUs.
- Review the MIT-licensed source and build it in a controlled environment if required
- Run the model server on the isolated network
- Point summaries at it in the app's AI settings
- Restrict access: Ollama has no authentication by default
What holdsAudio stays on the device. Transcript text stays on your network.
Posture 03 Connected
Your cloud tenancy, your keys.
Where the classification allows, summaries run in your organisation's own cloud account, such as AWS Bedrock or your OpenAI or Anthropic agreement, or through your organisation's gateway so desktops never hold provider keys.
- Choose a provider approved for the information involved
- Configure it centrally through your organisation
- Keep transcription on the device (the default)
What holdsStenoAI runs no server that receives meeting content.
Use in any classified space needs your accrediting authority’s approval of the device and the audio capture. Accreditation is a function of your own deployment and authorisation process, not the app alone.