Backchannel v0.3.7

Backchannel v0.3.7

The analysis agents can now run against a self-hosted OpenAI-compatible
server instead of a cloud provider. Paired with local ONNX transcription,
a deployment can run end to end on your own hardware with no API key from
anyone.

Highlights

• Self-hosted text models. A new openai-compatible entry in the model
  registry targets any OpenAI-shaped chat server: Ollama
  (http://localhost:11434/v1), LM Studio (http://localhost:1234/v1),
  vLLM, or LiteLLM. Set its base URL and model id in Admin -> API Keys and
  select it per agent like any other model.
• No credential required. Local servers usually want none, so this
  provider does not require an API key and sends no authorization header
  at all -- an empty bearer token breaks some servers rather than being
  ignored. Set OPENAI_COMPATIBLE_API_KEY only if your endpoint expects one.
• Layered configuration, unchanged defaults. The base URL resolves from
  the llm.openai_compatible.base_url app setting, then the
  OPENAI_BASE_URL environment variable, then the previous default. The
  app setting applies only to this provider, so a saved Ollama URL can
  never redirect a genuine OpenAI model. With nothing configured, every
  existing deployment behaves exactly as before.

Known limitation

Privacy First still switches the analysis agents off. Its gate recognizes
only the local ONNX transcription models, and an OpenAI-compatible URL can
point anywhere, so honoring it there needs a decision about whether to
require loopback. A fully local setup is configured through the endpoint
setting today, not by enabling that switch.

Not in this release

The WebSocket audio-handler refactor (ordered diarization worker, gateway
isolation, segment lifecycle serialization) is complete and reviewed on its
branch but is held back pending live dual-track call validation. It will
ship in a following release.

Upgrading

No migration or configuration change is required. Existing installs are
unaffected until an OpenAI-compatible base URL is set.

Backchannel v0.3.6

Backchannel v0.3.6

This hotfix completes OpenAI support across every agent and adds OpenAI
batch transcription options.

Highlights

• The Briefing trio (Meeting Lens, Discovery Lens, Arbiter) and the
  Strategic Signals agent now route by the selected model's provider.
  Previously they always called Gemini, so configuring them with an
  OpenAI model (for example gpt-5.6-luna) failed after the call with a
  Gemini "model not found" error. All text agents now work with any
  registry text model.
• Briefing failures now persist an actionable message (what happened and
  the remedy) instead of a raw provider error dump.
• Batch transcription gains OpenAI options: GPT Audio 1.5 and the
  cost-efficient GPT Audio Mini, transcribing diarized segment audio
  through OpenAI Chat Completions audio input. OpenAI's GPT-5.6 text
  models do not accept audio input, so they are deliberately not batch
  options. Both new models are key-gated and priced in Admin -> About.

Downloads

Executables are delivered through the authenticated Backchannel portal at
https://downloads.backchannel.page/. An approved Backchannel account is
required; no GitHub identity or repository membership is required. This
GitHub release retains the public source tag and notes without executable
files.

• Backchannel-windows-x64.zip - Windows 10 or later, x64
• Backchannel-macos-arm64.zip - macOS on Apple Silicon
• Backchannel-linux-x64.tar.gz - portable Linux x64 bundle

All desktop bundles are unsigned. Docker Compose users receive the same code
by rebuilding from the v0.3.6 source tag.

Backchannel v0.3.5

Backchannel v0.3.5

This release makes provider failures actionable everywhere they surface,
refreshes the OpenAI model lineup, and adds cost visibility for completed
calls.

Highlights

• Chat and Enhance Insights no longer fail with a bare "Internal Server
  Error" when a model provider hits its quota, spending cap, or rejects an
  API key. Errors now say what happened and how to fix it (raise the cap,
  or switch the model or provider in Admin), and a failed Enhance run
  shows its reason with a one-click "Retry failed batches" control.
• The post-call summary banner now anchors its numbers: "Final analysis
  pass: 3 new insights, 7 insights updated - 23 insights total for this
  session."
• All confirmations and notices use the app's themed dialogs; the raw
  browser "localhost says" popups are gone.
• The OpenAI lineup is current, including the GPT-5.6 family
  (gpt-5.6-sol, gpt-5.6-terra, gpt-5.6-luna). OpenAI transcription models
  (gpt-4o-transcribe, gpt-4o-mini-transcribe) can be selected as the
  batch transcription model once a working OpenAI API key is added.
• Admin -> About gains a Models & Pricing table served by the app, and
  the post-call Tokens tab shows estimated cost per model and per session
  at standard text rates (prices as of July 23, 2026).

Downloads

Executables are delivered through the authenticated Backchannel portal at
https://downloads.backchannel.page/. An approved Backchannel account is
required; no GitHub identity or repository membership is required. This
GitHub release retains the public source tag and notes without executable
files.

• Backchannel-windows-x64.zip - Windows 10 or later, x64
• Backchannel-macos-arm64.zip - macOS on Apple Silicon
• Backchannel-linux-x64.tar.gz - portable Linux x64 bundle

All desktop bundles are unsigned. Docker Compose users receive the same code
by rebuilding from the v0.3.5 source tag.

Backchannel v0.3.4

Backchannel v0.3.4

This hotfix release repairs starting a call, which was broken in v0.3.1
through v0.3.3.

Highlights

• Pressing Start Call on a new session failed silently in v0.3.1-v0.3.3:
  the request that marks the session active hit a server error, so the
  call never began. The server error is fixed and starting a call works
  again for new and existing sessions.
• No data was affected. Sessions created while the bug was present work
  normally after upgrading, and resuming completed sessions was never
  broken.
• A regression test suite now covers the session state transitions that
  stamp call start and end times.

Downloads

Executables are delivered through the authenticated Backchannel portal at
https://downloads.backchannel.page/. An approved Backchannel account is
required; no GitHub identity or repository membership is required. This
GitHub release retains the public source tag and notes without executable
files.

• Backchannel-windows-x64.zip - Windows 10 or later, x64
• Backchannel-macos-arm64.zip - macOS on Apple Silicon
• Backchannel-linux-x64.tar.gz - portable Linux x64 bundle

All desktop bundles are unsigned. Docker Compose users receive the same code
by rebuilding from the v0.3.4 source tag.

Backchannel v0.3.3

Backchannel v0.3.3

This release introduces a standalone Strategic Signals agent for live calls,
moves the full briefing pipeline to post-call or on-demand runs, and fixes a
bug that could permanently break the diarization diagnostics card.

Highlights

• A new standalone Strategic Signals agent keeps the live action cards
  current with a single model call per cycle (default every 45 seconds,
  configurable in Administration -> Agents). Evidence links and upvotes on
  its cards are preserved across refreshes.
• The briefing trio (meeting lens, discovery lens, and arbiter) now runs
  only after a session ends or on demand from post-call review. Live calls
  spend less, and the deliberate dual-lens briefing remains unchanged where
  it matters.
• Recording a mic benchmark on a machine without the optional Sortformer
  stack no longer stores an unusable result that left the Diarization
  Capability card stuck on "Unable to load diagnostics". Existing installs
  that hit the bug recover automatically after upgrading.
• Revoked recipient accounts can be reactivated from the operator console
  instead of remaining permanently locked out.
• The public landing page received a design polish pass, and the published
  agent documentation now matches the shipped agent crew and triggers.

Downloads

Executables are delivered through the authenticated Backchannel portal at
https://downloads.backchannel.page/. An approved Backchannel account is
required; no GitHub identity or repository membership is required. This
GitHub release retains the public source tag and notes without executable
files.

• Backchannel-windows-x64.zip - Windows 10 or later, x64
• Backchannel-macos-arm64.zip - macOS on Apple Silicon
• Backchannel-linux-x64.tar.gz - portable Linux x64 bundle

All desktop bundles are unsigned. Docker Compose users receive the same code
by rebuilding from the v0.3.3 source tag.

Backchannel v0.3.2

Backchannel v0.3.2

This release bundles ffmpeg with the Windows and Linux desktop builds so
voice calibration, mic benchmarks, and compressed audio imports work out of
the box, improves audio-conversion error messages, and cleans up two
confusing Administration labels.

Highlights

• The Windows and Linux desktop bundles now ship a self-contained LGPL
  ffmpeg, resolved automatically by the app. Recording a voice profile,
  running a mic benchmark, and importing MP3, M4A, or WebM audio no longer
  require installing ffmpeg separately. The macOS bundle still uses a
  system ffmpeg on PATH.
• Audio conversion failures now explain themselves: a missing ffmpeg
  produces a clear "install FFmpeg and make sure it is on PATH" message,
  and decode failures surface ffmpeg's actual error instead of a raw
  [WinError 2] system error.
• The Diarization Capability card no longer shows a status tag that could
  read as "diarization is unavailable" when only the optional Enhanced
  (Sortformer) mode was locked behind a benchmark.
• First-run setup describes the Cloud AI path accurately for both
  providers: a Google or OpenAI key powers the analysis agents and live
  captions, while saved transcripts come from Gemini or the built-in local
  transcription models.
• Startup is resilient to unrelated or legacy keys in a local .env file
  instead of failing validation.

Downloads

Executables are delivered through the authenticated Backchannel portal at
https://downloads.backchannel.page/. An approved Backchannel account is
required; no GitHub identity or repository membership is required. This
GitHub release retains the public source tag and notes without executable
files.

• Backchannel-windows-x64.zip - Windows 10 or later, x64
• Backchannel-macos-arm64.zip - macOS on Apple Silicon
• Backchannel-linux-x64.tar.gz - portable Linux x64 bundle

All desktop bundles are unsigned. Docker Compose users receive the same code
by rebuilding from the v0.3.2 source tag.

Backchannel v0.3.1

Backchannel v0.3.1

This release makes first-run setup genuinely guided, honors the Windows
default browser for the desktop app window, makes session groups safely
deletable, and adds observable, retryable speaker-revalidation batching.

Highlights

• Guided first-run provider onboarding: choose Google Gemini, OpenAI, or
  Privacy First local mode; follow the new step-by-step API key guide with
  screenshots and in-app "Get a key" links; readiness checks now verify the
  agents' active models against actually-usable credentials, so an
  OpenAI-only key no longer falsely completes setup while Gemini-default
  agents are unusable.
• The Windows desktop app window now opens in your default browser
  (resolved through the system's HTTP association) instead of preferring
  hardcoded browsers, with a standard browser tab as the fallback.
• Session groups are deletable from the sidebar via an accessible,
  confirmed control; sessions are preserved and simply ungrouped, never
  deleted with the group.
• Speaker-revalidation runs after speaker corrections are now batched with
  visible progress, honest per-batch outcome reporting, and retry for only
  the failed batches.

Downloads

Executables are delivered through the authenticated Backchannel portal at
https://downloads.backchannel.page/. An approved Backchannel account is
required; no GitHub identity or repository membership is required. This
GitHub release retains the public source tag and notes without executable
files.

• Backchannel-windows-x64.zip - Windows 10 or later, x64
• Backchannel-macos-arm64.zip - macOS on Apple Silicon
• Backchannel-linux-x64.tar.gz - portable Linux x64 bundle

All desktop bundles are unsigned. Docker Compose users receive the same code
by rebuilding from the v0.3.1 source tag. Note: voice calibration requires
ffmpeg on PATH when running desktop bundles.

Backchannel v0.3.0

Backchannel v0.3.0

This release adds local voice enrollment for reliable mic-only speaker
identity, flexible end-of-call analysis control, per-call token usage
tracking, transcription readiness safeguards, split-track retranscription
continuity, and in-app version/release-notes surfaces with first-run
onboarding.

Highlights

• Enroll the local user's voice once (Administration -> Transcription &
  Audio); mic-only calls then map that voice to the sole local user. Only an
  encrypted embedding is stored; calibration audio is never persisted.
• Split End Call button: full pipeline by default, "End without briefing" to
  skip briefing synthesis and opportunity matching. Unintentional
  disconnects finalize the session with zero analysis LLM calls, and
  post-call review offers on-demand briefing generation.
• Per-call token usage is captured at every backend LLM call site and shown
  in a post-call Tokens breakdown by source and model.
• Calls refuse to start with an actionable message when the selected batch
  transcription model has no usable credential; runtime transcription
  failures are counted and surfaced instead of completing silently empty.
• Split-track calls persist per-track audio provenance so retranscription
  preserves local/remote identity and speaker continuity across segments,
  ordered by source speech time.
• Post-call Speakers tab readability and accessibility fixes, with honest
  partial/error reporting and retry for speaker-aware revalidation.
• Administration gains an About tab (version + release notes); first launch
  shows a guided setup checklist; upgrades show a what's-new notice.
• Diarization: short startup fragments no longer seed permanent speaker
  profiles; the 0.68 similarity threshold is retained with documented
  noise-floor evidence.

Downloads

Executables are delivered through the authenticated Backchannel portal at
https://downloads.backchannel.page/. An approved Backchannel account is
required; no GitHub identity or repository membership is required. This
GitHub release retains the public source tag and notes without executable
files.

• Backchannel-windows-x64.zip - Windows 10 or later, x64
• Backchannel-macos-arm64.zip - macOS on Apple Silicon
• Backchannel-linux-x64.tar.gz - portable Linux x64 bundle

All desktop bundles are unsigned. Docker Compose users receive the same code
by rebuilding from the v0.3.0 source tag. Note: voice calibration requires
ffmpeg on PATH when running desktop bundles.

Backchannel v0.2.5

Backchannel v0.2.5

This patch release gives Backchannel a dedicated desktop-app window when a
supported Chromium browser is installed and updates the default Gemini models
used for meeting analysis and batch transcription.

Highlights

• Open Backchannel in a dedicated app window with its own taskbar or Dock
  presence when Chrome, Edge, or Chromium is installed; otherwise use the
  default browser.
• Reserve a stable local port and verify each running instance before launch so
  shortcuts and app-window behavior remain reliable.
• Advertise an installable web-app manifest with purpose-specific application
  icons.
• Add Gemini 3.6 Flash and Gemini 3.5 Flash-Lite as selectable models.
• Default consolidated analysis, opportunity analysis, and the three briefing
  lenses to Gemini 3.6 Flash.
• Default objection handling and batch transcription to Gemini 3.5 Flash-Lite.
• Apply the new defaults once to existing installations while preserving any
  choices users make after upgrading.

Downloads

Executables are delivered through the authenticated Backchannel portal at
https://downloads.backchannel.page/. An approved Backchannel account is
required; no GitHub identity or repository membership is required. This GitHub
release retains the public source tag and notes without executable files.

• Backchannel-windows-x64-v0.2.5.zip - Windows 10 or later, x64
• Backchannel-macos-arm64-v0.2.5.zip - macOS on Apple Silicon
• Backchannel-linux-x64-v0.2.5.tar.gz - portable Linux x64 bundle

All desktop bundles are unsigned. Docker Compose users receive the same code by
rebuilding from the v0.2.5 source tag.

Backchannel v0.2.4

Backchannel v0.2.4

This patch release improves two-speaker diarization, local transcription first
use, and the memory profile of the default lightweight audio path.

Highlights

• Split internally mixed long turns with short coherence windows while keeping
  the established 0.68 speaker-similarity threshold unchanged.
• Keep coherence-group assignment non-enrolling so an ambiguous transition
  cannot create extra speaker profiles or consume registered remote slots.
• Allow Whisper and Parakeet local ASR models to download on first use without
  an API key, then reuse the cached models offline.
• Skip optional Sortformer probing for normal live calls and audio imports that
  use lightweight diarization, and launch the backend directly to reduce
  avoidable model and process overhead.

Split-track labeling

On a live split-track call, the sole configured local user is assigned only to
the physical microphone. Every voice arriving through shared system audio is
diarized normally as a remote participant. This also means that a recording
played from another tab keeps both recorded speakers in the remote clusters,
even when one recorded voice belongs to the person operating Backchannel.

Downloads

Executables are delivered through the authenticated Backchannel portal at
https://downloads.backchannel.page/. An approved Backchannel account is
required; no GitHub identity or repository membership is required. This GitHub
release retains the public source tag and notes without executable files.

• Backchannel-windows-x64.zip - Windows 10 or later, x64
• Backchannel-macos-arm64.zip - macOS on Apple Silicon
• Backchannel-linux-x64.tar.gz - portable Linux x64 bundle

All desktop bundles are unsigned. Docker Compose users receive the same code by
rebuilding from the v0.2.4 source tag.