Echo’s production path is designed around local processing:
- Dictation recognition runs locally through Core ML.
- Meeting intelligence runs locally through MLX.
- Deterministic vocabulary and spoken-command rules remain on-device.
- The production route does not require a transcription API.
- Completed meetings retain transcripts and notes while source-audio retention is minimized.
- Obsidian export is one-way and does not write source audio to the vault.
“Local-first” describes the implemented inference path. It does not mean that a macOS app needs no permissions or release security work.
Depending on the selected workflow, Echo may require:
- Microphone access for dictation and live recording
- Accessibility and Input Monitoring for the global shortcut and active-app insertion
- Screen Recording permission for system-audio meeting capture
Permissions are requested for the corresponding workflow and are not a substitute for clear user consent when recording a meeting.
Recording laws and workplace policies vary by jurisdiction. Echo’s user is responsible for notifying participants and obtaining any consent required for the meeting and location involved.
The private development repository must remain private. Its history contains material that should not be exposed even if the current working tree is reviewed.
This case-study repository intentionally excludes:
- Source code and proprietary product logic
- Git history from the product repository
- Credentials, signing identities, and notarization records
- Model weights
- Private transcripts and benchmark references
- Tester information and unreleased features
A build working on the development Mac is not automatically a public macOS release. Distribution to other Macs requires the appropriate Developer ID signature, hardened-runtime configuration, notarization, ticket stapling, and clean-Mac Gatekeeper validation.
This repository does not distribute an Echo binary and should not be interpreted as a public product release.