Toki connects a deliberately small chain from an authorized Discord start to a durable, permission-scoped transcript.
The host starts from Discord, the configured room and linked destination are resolved, and a visible transcription notice arrives before capture.
Discord voice is decoded into bounded per-speaker frames and streamed to speech recognition. Cleanup closes receiver and provider resources on stop or failure.
Final English transcript events are persisted idempotently with meeting identity, stable Discord attribution, timestamps, confidence, and provider metadata.
Toki keeps requester access, destination disclosure, and external-action authority separate. Missing or uncertain policy does not become permission.
The deterministic application owns identity, lifecycle, persistence, permission, and action state. A model may help with language later; it does not own authority.
Access follows identity. Authority follows explicit policy.
Current capture depends on three explicit providers. GitHub recall and approved issue creation remain later roadmap work.