Voices become memory
Call Scribe records architecture calls and writes them down — diarized transcripts, decisions, action items, and a Codex-ready handoff, filed in your repo as durable Markdown memory.
The best architecture discussion happens in a voice channel — and then it evaporates. We built the notebook that doesn't forget.
A Discord voice channel, a phone line, a recording. Call Scribe joins, records whoever speaks, and stops when the channel empties — phone handoffs included. The raw audio becomes the source of truth. Oversized recordings are pre-split into ten-minute mono chunks before transcription, so a 43-minute architecture call never times out.
The analysis pass classifies every beat: Decision, Constraint, Open Question, Owner. Speaker turns render as diarized Markdown — who said what, in order. The taxonomy is opinionated on purpose: it is the same structure your repo can act on, the same headings a Codex task file consumes.
The manuscript lands in your tree — transcript, architecture brief, analysis, and a Codex task file. Not a silo, not a dashboard: docs/meetings in the repo where the implementation will happen. Self-hosted with Docker Compose or on Kubernetes; the audio and the writing never leave your infrastructure.
Every meeting ends with a task file your coding agent consumes directly — codex-task.md. The decision becomes the diff, the diff becomes the code, and the code is what the next architecture call reviews. That is the loop that makes the call count.
Open source · Apache-2.0
Self-hosted with Docker Compose or on Kubernetes. The drawings, transcripts, and decisions stay in your repo — yours to keep.
