Explicit action policy
A microphone permission or connected account is not approval to send, purchase, delete, or change a device. Consequential actions need an explicit review step.
Technology · Proposed architecture
Natural voice. Deliberate reasoning.
Two complementary layers, connected by an application that keeps the task, permissions, and result in view. OpenAI handles live conversation; Claude is planned to do the deeper reasoning behind the work.
Claude at the core of the task
From a request to an informed next step.
The proposed Claude layer does more than generate a reply. It breaks down goals, decides which permitted tool fits, evaluates what comes back, and identifies context worth keeping.
The website demonstrates the product concept. No live voice, model orchestration, tool execution, or persistent agent memory is running here.
The intended loop
A clear path from voice to verified result.
01
You speak. The voice layer listens.
OpenAI GPT-Live 1 is the proposed full-duplex voice layer, with interruptions and a natural spoken response. The application owns the transcript and task context.
02
The application carries the task.
An orchestrator maintains the goal, conversation, access boundaries, usage budget, and pending approvals. Client delegation is an application integration, not an automatic handoff between vendors.
03
Claude makes a plan.
Claude Sonnet 5.5 is the proposed default for planning, tool selection, interpreting results, and refining context. Opus 5.5 may be used for harder tasks within a defined budget.
04
A permitted tool does the work.
The executor checks access and action policy, obtains the required approval, and runs a bounded task through an API or local companion. Screenshots belong in this separate reasoning and tool path.
05
The result is checked and explained.
The application validates the tool result, records the outcome, and returns grounded context for a spoken reply. Failed and partial actions remain visible.
06
Useful memory is proposed, not assumed.
Claude may suggest a preference or fact to retain. Planned controls require your approval and let you inspect, correct, and delete saved memories.
Designed into the foundation
Context, control, and a cost ceiling.
Bounded execution
Tasks need limits on tool steps, time, tokens, and concurrency. Interruptions should preserve pending work without turning an ambiguous transcript into authorization.
Visible usage
Live voice minutes and task credits are separate budgets. Long tasks should show an estimate for approval, and stop safely at the allowance.
A concrete data flow
Local access does not mean local-only processing.
A planned local companion can keep execution on your device. The proposed voice and reasoning services still process selected data in the cloud.
Audio and transcripts would be sent to OpenAI’s configured voice service. Selected task context, tool results, and any permitted screenshots may go to the Claude API. The application must govern retention, permissions, memory, and deletion.
API credentials belong on a trusted server or local companion. Browser voice requires a microphone and HTTPS or localhost. Availability, latency, and cost must be validated in a real pilot before launch.
In development
Help us test a useful first workflow.
Tell us the work you would like Vynce to help with. Join the waitlist for early-access updates.