backchannel / site

# Backchannel

> Backchannel is a self-hosted, open-source (MIT) AI meeting assistant. It
> captures meeting audio in the browser, builds a live speaker-attributed
> transcript using local voice activity detection and speaker embeddings,
> and runs configurable AI agents (Google Gemini or OpenAI, with optional
> fully local ONNX transcription) that surface questions, objections,
> opportunities, and action items while the meeting is still happening.
> You can also ask the running call a question and get an answer from the
> conversation so far, without stopping the recording. A PII Shield can
> replace names, companies and contact details with tokens before anything
> is stored or shown to a model, local or cloud, so the models read the
> conversation without reading the people in it.

Backchannel is free to run: there is no hosted service and no pricing
tiers, and self-hosting with Docker Compose (React frontend, FastAPI
backend, PostgreSQL) requires no account of any kind. The prebuilt desktop
executables for Windows, macOS, and Linux are a free direct download from
the public download portal; no account is needed. Bring your own Gemini or OpenAI API key, or transcribe
fully offline with local ONNX Whisper/Parakeet models. Because it is
self-hosted, meeting audio and transcripts never leave your infrastructure
except for the model API calls you configure.

Key facts:
- License: MIT, source at https://github.com/talberthoule/backchannel
- Category: real-time meeting analysis / AI meeting assistant / meeting copilot
- Desktop release: v0.6.6 for Windows x64, macOS Apple Silicon, and Linux x64, downloaded freely at https://downloads.backchannel.page/; no account, GitHub identity, or repository membership is required. From v0.4.0 the desktop app updates itself in place with a verified signature
- Self-hosted alternative to cloud meeting assistants such as Otter.ai,
  Fireflies.ai, and Granola
- Stack: React 18 + TypeScript frontend, FastAPI backend, PostgreSQL 16
- Diarization: Silero VAD + WeSpeaker ResNet152 speaker embeddings (local, ONNX)
- Transcription: Gemini Flash, OpenAI batch audio models, or fully offline
  ONNX Whisper/Parakeet; interim live captions come from Gemini Live or
  OpenAI Realtime by default, or run on-device with the local Parakeet
  captioner
- Text analysis: Google, OpenAI, or any number of self-hosted
  OpenAI-compatible endpoints (Ollama, LM Studio, vLLM, LiteLLM), with
  every served model listed by name in the pickers -- so a deployment can
  run end to end with no cloud key. Privacy First judges the destination,
  so an on-prem endpoint (loopback or LAN) keeps the analysis agents
  running with the switch on while cloud providers stay blocked.
- Agents: ten seeded agents -- an audio gateway, six live agents
  (consolidated analyst every 40s, objection handler every 10s over the last
  90s, synthesizer, opportunity specialist, strategic signals every 45s, and
  a transcript refiner every 45s that is off by default), and a three-agent
  post-call briefing pass (meeting lens, discovery lens, arbiter) -- each
  with a configurable model, prompt, and trigger
- PII Shield: personal data is replaced with tokens such as [PERSON_1] at
  the moment a transcript line, directive, document excerpt, session name or
  speaker name is written, so every agent, briefing, chat and Ask works from
  tokenized text and the database itself holds only tokens. It covers
  people's names, company names, email addresses, phone numbers, payment
  card and national-identifier numbers, IP addresses and street addresses.
  Real values live in a vault encrypted under the same master key that
  protects provider keys, and are restored only on the operator's own
  screen and in exports explicitly asked for. Detection is entirely local:
  patterns for structured identifiers, the session's speaker roster, a
  user-maintained protected-terms list, and an on-device name-recognition
  model (Xenova/bert-base-NER, about 110 MB) that downloads once and then
  runs offline. No detection ever makes a cloud call. Because audio cannot
  be tokenized, the shield holds transcription to a local model and turns
  off cloud live captions; cloud text models stay usable because they
  receive tokens only. An optional prompt log records every outbound prompt
  exactly as it left, badged "tokens only" or "blocked", so the claim is
  checkable rather than trusted; while the shield is on, a prompt still
  carrying a vault value is refused before it is sent
- Transcript refiner: an optional agent that sends the tokenized transcript
  to any text model, local or cloud, to fix punctuation, casing, sentence
  boundaries and mishearings, closing the quality gap of local-only
  transcription. A rewrite is kept only if it carries exactly the original
  tokens, so no name reaches the model
- Cost transparency: a per-session Tokens tab breaks usage down by source
  and by model, pricing cached prompt tokens and audio tokens at their own
  published rates rather than the text rate, and the post-call Overview
  carries the resulting estimate
- Ask during the call: the call's command bar answers a typed question from
  the live transcript, the insights raised so far, strategic signals,
  directives, and stored document summaries, without stopping the recording.
  The answer is saved as a starred `asked` insight and exported with the
  rest. It is not an agent: nothing schedules it and asking does not steer
  the running agents. The answering model is chosen in the bar and may be a
  self-hosted one
- Strategic signals persist: signals raised during a call are kept with a
  per-signal count and first/last sighting, readable live and in the
  post-call briefing, rather than being overwritten by the next cycle
- Nothing is chosen for you at first run: a fresh install seeds no cloud
  model, every agent starts unselected, models are grouped by Google,
  OpenAI, and local, and one role-appropriate recommendation is marked in
  each provider that is actually available
- Works with any meeting, digital or in person: captures mic plus tab/system
  audio in the browser, so Zoom, Meet, and Teams need no bot or integration,
  and a conference-room conversation needs only the microphone

## Privacy

- [Private meeting AI](https://backchannel.page/private-meeting-ai/): the full account
  of the PII Shield -- what is tokenized and when, the layered on-device detection,
  the encrypted vault, the three places a value is decoded, why audio is held to
  local models while cloud text models stay usable, how to verify the claim with the
  outbound prompt log, and the limits stated plainly (off by default, detection is
  not perfect, tokenization is reversible by design, and it governs what models read
  rather than who can reach the machine)
- [Three degrees of local](https://backchannel.page/blog/three-degrees-of-local/): where
  each component can run, from local diarization through fully self-hosted analysis

## How it works

- [Analyst lenses](https://backchannel.page/analyst-lenses/): how the Consolidated
  Analyst reads the call. Every 40 seconds one request carries the latest 30
  transcript lines and one prompt section per enabled lens; the reply uses one
  fixed shape, each finding tagged with its type and lens. Covers custom lens types,
  the Opportunity lens being left out of meeting types without offering matching,
  and the trade-off that all lenses share one model, cycle, and window

## Docs

- [Download portal](https://downloads.backchannel.page/): public Windows, macOS, and Linux executable downloads; no account required
- [Desktop release v0.6.6](https://backchannel.page/releases/v0.6.6/): public asset inventory, release notes, install notes, and first-run guidance
- [All releases](https://backchannel.page/releases/): version history with publish dates and per-release notes
- [Full documentation as one file](https://backchannel.page/llms-full.txt): every docs page concatenated as plain markdown for agents
- [Quickstart](https://backchannel.page/docs/quickstart/): Docker Compose setup, local development, migrations, tests
- [Architecture](https://backchannel.page/docs/architecture/): the live call path end to end
- [Agent System](https://backchannel.page/docs/agents/): each agent's trigger, purpose, and configuration
- [Audio Pipeline](https://backchannel.page/docs/audio-pipeline/): capture format, VAD/diarization, transcription routing
- [WebSocket Protocol](https://backchannel.page/docs/websocket-protocol/): binary audio framing and message types
- [REST API](https://backchannel.page/docs/rest-api/): endpoint reference
- [Configuration](https://backchannel.page/docs/configuration/): settings, environment variables, model registry
- [Deployment](https://backchannel.page/docs/deployment/): compose services, GPU support, startup behavior

## Writing

- [Writing index](https://backchannel.page/blog/): dated, vendor-sourced notes on the AI meeting-assistant category
- [The free AI notetaker is genuinely free. Here is why I self-host anyway.](https://backchannel.page/blog/self-hosted-ai-notetaker-honest-comparison/) (2026-07-28): sourced competitive research against Fathom, Otter, Read AI, tl;dv, and Avoma that opens with the concessions -- Fathom's free tier really is unlimited, bot-free capture is no longer a differentiator, and real-time in-call AI shipped at Otter and Zoom in July 2026 -- then argues from what survives: where audio is processed, whether a vendor trains on conversations, tactical generated responses vs descriptive live summaries, configurability, and readable source.
- [Three degrees of local: what "self-hosted" actually means in a meeting assistant](https://backchannel.page/blog/three-degrees-of-local/) (2026-07-28): architecture walk-through showing locality as a three-position spectrum chosen per component. Diarization is always local (Silero VAD plus WeSpeaker ResNet152, ONNX); degree one pairs local diarization with cloud transcription on your own key; degree two runs transcription and live captions fully on-device; degree three points the analysis agents at a self-hosted OpenAI-compatible endpoint (Ollama, LM Studio, vLLM, LiteLLM) so no packet leaves your network. Every claim cited to a file path.
- [Live meeting AI arrived in one week, and all of it is Enterprise-gated](https://backchannel.page/blog/live-meeting-ai-enterprise-gate/) (2026-07-25): Otter Live Assist (2026-07-21, Otter Enterprise) and Zoom Revenue Accelerator Sales Assist (2026-07-22, Essentials from $66/user/mo and Premium from $99.99/user/mo annual, Essentials dated August) both shipped real-time in-call guidance within 48 hours. Compares the three mechanisms now in market: Zoom pre-configures guidance per topic category via admin setup, Otter grounds it in playbooks and SOPs you supply, and Backchannel generates it from the trailing transcript window. Concedes that pre-authored guidance is more consistent and reviewable.

## Compare

- [Open-source Otter.ai alternative](https://backchannel.page/otter-alternative/): self-hosted transcripts and insights without minute caps
- [Open-source Granola alternative](https://backchannel.page/granola-alternative/): bot-free capture, self-hosted and real-time
- [Fireflies.ai alternative without the bot](https://backchannel.page/fireflies-alternative/): live assistance on your own hardware
- [Backchannel vs Meetily](https://backchannel.page/vs-meetily/): two open-source meeting assistants compared
- [Open-source meeting assistants](https://backchannel.page/open-source-meeting-assistants/): the self-hosted landscape (Meetily, Anarlog, Amurex, Vibe, Buzz, Backchannel) compared honestly, including when to pick something else
- [Backchannel vs Anarlog, formerly Hyprnote](https://backchannel.page/vs-anarlog/): a local-first Mac notetaker vs a live agent server; both keep audio local, only one acts during the call
- [Fathom alternative](https://backchannel.page/fathom-alternative/): what a generous free tier does and does not cover
- [Read AI alternative](https://backchannel.page/read-ai-alternative/): meeting intelligence without the auto-join governance problem
- [Gong and Backchannel](https://backchannel.page/gong-and-backchannel/): not a Gong replacement -- where a live seller-side layer fits under revenue intelligence
- [Backchannel vs Clari Copilot](https://backchannel.page/vs-clari-copilot/): responses generated against the objection actually raised vs pre-written battlecards
- [Teams Premium alternative](https://backchannel.page/teams-premium-alternative/): one assistant across every meeting platform instead of one per platform
- [Plaud alternative](https://backchannel.page/plaud-alternative/): bot-free capture without buying hardware or sending audio to a vendor cloud

## Source

- [GitHub repository](https://github.com/talberthoule/backchannel): MIT-licensed source code and README