Backchannel Documentation
Backchannel is a real-time meeting analysis app. A React frontend captures microphone (and optionally tab/system) audio, streams it to a FastAPI backend over WebSocket, and the backend produces a speaker-attributed transcript while a set of provider-routed AI agents surface questions, observations, opportunities, objections, and action items as the conversation happens. The call’s command bar also answers a question put to the conversation so far, without interrupting it (Agent System).
This folder is the deeper technical reference behind the top-level README. For how Backchannel differs from the other meeting assistants people evaluate, the public site’s comparison hub indexes every open-source and commercial comparison.
For desktop operations, Releasing and the private R2 manifests
are authoritative. GitHub releases retain public source tags and notes without
executable files; the executables themselves are free public downloads from
the download portal. The Cloudflare Access-protected operator console remains
for the account machinery behind the deployed desktop updater’s grant flow: it
separates Early access request/consent approval and rejection, Users
identity/security commands, and Authorization Latest/explicit-version grants. Authorization is
stored in release_access_policies plus release_account_versions; the old
/api/admin/access/* routes are removed. Recipient accounts, grants, sessions,
and access events live in D1, not the local application’s PostgreSQL database.
Contents
Section titled “Contents”| Page | What it covers |
|---|---|
| Quickstart | Running with Docker Compose, local development, database migrations, tests |
| Getting API Keys | Creating a Google Gemini or OpenAI key and connecting it under Admin -> Connections |
| Architecture | The live call path end to end, frontend structure, backend key files |
| Agent System | The orchestrator, each agent’s trigger and purpose, configuration and overrides |
| Audio Pipeline | Capture format, VAD and diarization, batch transcription routing, audio storage |
| WebSocket Protocol | Binary audio framing and every JSON message type on /ws/{session_id} |
| REST API | Endpoint reference for all routers, grouped by resource |
| Configuration | Settings, environment variables, encrypted credentials, the model registry |
| Deployment | Docker Compose plus the ordered D1, R2, Worker, Turnstile, and release-access production gate |
| Releasing | Authoritative private-R2 publication, migration, verification, and recovery procedure |
Reading order
Section titled “Reading order”If you are new to the codebase, read Quickstart to get the stack running, then Architecture for the mental model, then Agent System – most feature work touches one of those three areas. The protocol and API pages are references to consult as needed.
Release-access documentation changes must pass:
cd docs-sitenpm run test:release-accessnpm run test:migrationnpm run test:workernpm run test:adminnpm run test:downloadnpm run test:sitenode --test *.test.jsnpm run build