The BEFORE.Band founding batch is forming.Reserve yours
FROrder
Guide · External AI agents

What should an AI agent receive from a voice recorder?

A private recorder should deliver faithful words and their provenance, then leave summarization, reasoning, memory and action decisions to the external agent the user already chose.

BEFORE.Band · Published 16 July 2026 · 5 min read

A practical explanation of the boundary between BEFORE’s transcript-delivery responsibility and the downstream interpretation performed by ChatGPT, Claude Code or another user-controlled agent.

BEFORE.Band present during a conversation that gives a chosen AI personal context

Start with clean evidence

An external agent can only reason well when its source is faithful. BEFORE therefore produces readable transcript text while preserving capture time, language, timestamps, version history and processing provenance. V1 speaker keys remain null rather than being invented.

Cleanup may repair punctuation, spacing and obvious disfluencies only when meaning, attribution and evidence alignment remain unchanged. Uncertainty stays visible instead of being rewritten into confidence.

A transcript is not a task database

BEFORE does not turn speech into people, projects, tasks, decisions, questions, ideas, summaries, memories or candidate actions. Those are interpretations, not transcription outputs.

Keeping these objects out of the product avoids duplicating the language-model capabilities of the user’s own agent and makes it possible to inspect exactly what source material was delivered.

MCP delivers new material incrementally

An authorized agent polls for new transcripts with an opaque cursor. Repeating a cursor is safe and returns a stable progression. The agent can also retrieve one transcript by stable identifier.

MCP returns the clean transcript and provenance, never raw audio. It has no semantic search, identity-resolution or action tools in the transcript-delivery MVP.

The external agent owns interpretation

After retrieval, ChatGPT, Claude Code or another agent can summarize, extract, integrate memory or propose next steps according to the user’s instructions and that provider’s policies.

BEFORE does not provide this downstream agent and does not silently add transcripts to every conversation. Connection means future access is possible when the agent explicitly calls MCP.

A transcript is not authority

A recording may contain words such as “send the message” or “buy the ticket.” Preserving those words faithfully does not authorize any system to follow them.

Recorded speech remains untrusted source material. Any consequential action requires authorization in the external system that would perform it, separate from capture and transcript delivery.

Keep reading
What makes an AI voice recorder private?A concrete checklist for evaluating privacy in a wearable AI recorder—beyond a vague promise that data is secure.Should an AI wearable always listen?A practical comparison of intentional and ambient AI wearables, including the trade-offs in recall, privacy, bystander consent, data volume and everyday trust.The intentional AI wearable checklistA reusable, evidence-bounded framework for journalists, reviewers, builders and buyers who need to distinguish intentional capture from vague privacy marketing.How can you talk to your AI without taking out your phone?Yes—but “talk to your AI” can mean two different things. A watch or earbuds can support live back-and-forth. A one-press recorder can preserve a selected thought as a transcript your chosen AI can retrieve later, without turning your wrist into another screen.A reproducible test protocol for wearable transcript qualityThe scenarios, measures and reporting rules BEFORE.Band will use before publishing any transcript-quality result from a physical sample.

One press. One faithful transcript.

Reserve your band