Changelog
Everything that shipped to production, newest first. Desktop releases go out through the in-app updater.
New paid signups were getting a “welcome to Free” email moments before their upgrade landed. The welcome email now holds until checkout resolves, so it names the plan you actually bought. The template was rebuilt at the same time.
Re-transcribe any meeting with speaker detection so every line is attributed to who said it. It runs per meeting rather than globally, so you can spend the extra processing only on the calls where attribution matters.
Whisper can fall into a loop and repeat the same phrase for minutes on quiet or noisy audio. Reverbic now detects looped output, discards it, and retries that chunk at a higher temperature instead of shipping the garbage into your summary.
Share links now render the full meeting — summary, key points, decisions, action items, and transcript — server-side, so they load fast and preview properly when pasted into Slack or Linear.
The MCP server now enforces plan access server-side and honors a separate scope for transcript access, so a key can list meetings without being able to read what was said. Setup is paste-ready for Claude, Cursor, and Codex.
Browser recordings are captured as WebM, which cannot be seeked reliably. Reverbic now transcodes a playback copy to M4A once, so the player scrubs instantly instead of re-buffering from the start.
The CLI rendered raw ANSI escape codes on some terminals and mis-read its own token expiry, forcing needless re-logins. Both are fixed.
The multi-step signup wizard is gone. New accounts land on a real demo meeting they can explore immediately, with an inline getting-started checklist instead of a modal sequence to click through.
Routine hardening pass: dropped an over-permissive row-level security policy that exposed public meetings to anonymous reads, pinned the search path on the signup trigger, and started rejecting automated scanner probes at the edge.
Stability and packaging fixes across four releases, delivered through the in-app updater.
Ask questions across your meeting history in plain language — “what did we decide about the migration?” — and get an answer grounded in your transcripts, with the meetings it drew from. Shipped alongside a rebuilt landing page, recording page, and dashboard.
Every public page now self-canonicalizes and appears in the sitemap, so search engines stop collapsing distinct pages into one another.
Silent or near-silent audio could cause the transcriber to hallucinate plausible-sounding dialogue. Silence is now detected and skipped. Atmos was also scoped to coding work only, rather than trying to board every conversation.
The desktop app gained a real native dashboard window instead of bouncing you to the browser for everything beyond recording. Shipped across v2.0.29 – v2.0.32.
Long meetings are the ones you least want to lose. The pipeline now retries transient transcription failures, splits into 15-minute chunks, and keeps a cross-tab heartbeat so a closed tab no longer strands a job halfway through.
Hand someone the entire meeting — not a read-only link, but real access — through an invite-and-accept flow with an email notification.
The weekly digest toggle in settings now does something: a scheduled job sends a real summary of the week's meetings, decisions, and open action items.
PR generation is the expensive call in Atmos, so it now carries a per-organization daily limit shared across members. Atmos also got its own page explaining what it does.
Atmos turns a live conversation into an editable visual board of functions, tables, and endpoints — then grounds that board in your connected GitHub repository and drafts a real pull request. Nothing is committed until you confirm, and changes land on a branch for normal review.
Context Packs carry your stack, conventions, key files, and constraints into every spec. The system was rebuilt so multiple packs can be active at once, packs can be shared across your organization, and the active pack primes transcription vocabulary — so “pgvector” stops coming out as “PG factor.”
The desktop app records system audio on macOS and Windows, so calls in Slack, Meet, Teams, or anywhere else are captured without a bot joining as a participant. Slack Huddles can be recorded and transcribed automatically, with the summary posted back to the channel.
Recording is now durable: audio is written continuously, survives a crash or accidental navigation, and the raw file uploads even if processing later fails. A persistent recording bar lets you move around the app mid-recording.
Add domain vocabulary before processing so names, products, and acronyms come through correctly. Recordings and transcripts can now be downloaded directly, and shared pages show distilled specs.