VoiceGuard — Human-supervised voice operations

One shared page. Two kinds of judgment.

The agent finds the signal.
The human owns the action.

A live supervision cockpit where a browser agent can inspect, audit, and prepare—without ever dialing, speaking, or pausing on its own.

AUTONOMY BOUNDARY

✓Read and audit shared state

✓Draft guidance for review

×No dial, send, pause, or hang up

CALL STATE

LIVE

ROUND TRIP

812 mssynthetic

POLICY SIGNAL

REVIEW1 open item

HUMAN CONTROL

ENFORCEDtrusted click

SYNTHETIC SESSION VG-204

Live conversation

Jordan MillerFictional participant · verified test line

02:18

SIMULATED MEDIA STREAM

BROWSER AGENT WORKSPACE

Review queue

4 SITE TOOLS

POLICY REVIEW

Avoid an implied guarantee

The participant asked whether the next step is binding. Keep the reply factual and reserve commitments for a human follow-up.

AGENT DRAFT · NOT SENT

Operator guidance

Empty

0/240 · human review required

VISIBLE TOOL ACTIVITY

Agent trace

Shared

TOP-LEVEL IMPERATIVE API

Four narrow tools. Zero hidden authority.

Every WebMCP call changes—or reads—the same state the operator sees. Sensitive actions stop at a visible confirmation.

01 · READ

inspect_call_state

Current phase, metrics, transcript, and pending human actions.

02 · READ

audit_guardrails

Deterministic policy signals with untrusted transcript labeling.

03 · DRAFT

draft_operator_guidance

Fills the review box. It never touches the synthetic call.

04 · REQUEST

stage_pause_request

Raises a visible request. Only the human can confirm.

DEMO-ONLY NATIVE CLIENT · ?inspector=1

Chrome WebMCP execution trace

These controls discover registered tools with getTools() and invoke them with executeTool(). They are a transparent test client, not additional site tools.

Waiting for a native WebMCP call…