# Vynce architecture reference

This is a proposed architecture, not the current website's runtime. No live voice session, model orchestration, tool executor, or persistent agent memory is implemented by the website demo.

## Proposed task path

1. **OpenAI GPT-Live 1** handles live spoken conversation.
2. An **application orchestrator** maintains transcripts, conversation context, pending tasks, permissions, and cancellation state.
3. **Claude Sonnet 5.5** handles planning, tool selection, result interpretation, context refinement, and proposed memory changes.
4. **Permissioned tools** execute bounded actions after applicable approval checks.
5. The application validates results and returns a spoken response.

**Claude Opus 5.5** is an optional escalation for harder reasoning, subject to budget and latency limits. Sonnet is the proposed default. Cross-provider delegation requires application code; it is not an automatic built-in handoff.

The verified proposed voice model is `gpt-live-1`, not “GPT Live 1.5.” `gpt-realtime-1.5` is a separate model with a different cost structure. If that model is selected, the voice allowances must be recalculated from measured usage.

## Implementation requirements

- Microphone permission and a secure browser origin, plus trusted server-side credentials.
- Explicit tool scopes, bounded execution, cancellation, result checks, and approval for consequential actions.
- Application-owned conversation state across interruptions; ambiguous speech is not authorization.
- Only user-approved persistent memories, with inspection, correction, and deletion controls.
- Cost metering, per-task limits, concurrency controls, and a hard monthly budget.

Local tool access does not make model processing local-only: audio/transcripts would go to the configured OpenAI service, and selected context or screenshots may go to Claude. Publish concrete data flows and retention terms before launching these capabilities.

## Sources

- [GPT-Live 1](https://developers.openai.com/api/docs/models/gpt-live-1)
- [OpenAI client delegation](https://developers.openai.com/api/docs/guides/live-delegation)
- [Claude model overview](https://platform.claude.com/docs/en/models/overview)
- [Claude pricing](https://platform.claude.com/docs/en/about-claude/pricing)

See [the technology page](/technology), [plan accounting](/reference/plans.md), and [privacy](/reference/privacy.md).