Incremental polling
An authorized external agent requests only transcripts created after its last opaque cursor. Repeating the same cursor is safe and idempotent.
MCP delivery
Faithful transcripts in. Your agent takes it from there.
BEFORE ends at a clean transcript with evidence. Your external agent periodically polls the MCP service, retrieves new material and decides what it means according to your instructions.
Exact retrieval
An agent can fetch a specific transcript by identifier and receive its current version, capture time, language, timed segments and provenance. V1 speaker keys are null rather than invented.
Transcript-only boundary
MCP never returns raw audio or BEFORE-generated tasks, decisions, identities, summaries, projects, people, ideas or actions.
Your agent decides
ChatGPT, Claude Code or another compatible agent decides whether to summarize, reason, remember or act under your own instructions. BEFORE does not provide that agent.
Two MCP operations, one narrow responsibility
Poll for new transcripts with an opaque cursor, or retrieve one transcript by its ID. BEFORE performs no semantic search, identity resolution, summarization or action loop.