Converses, searches, compares, and prepares an inert draft.
Agent-native automotive retail
A 24/7 digital sales channel — not a chatbot
The agent understands the buyer and calls explicit tools. A public deterministic policy protects dealership margin. Only a person can approve evaluation, and the final commercial call remains human-controlled.
Produces the same decision for the same approved input.
Opens the evaluation gate and makes the final call.
Checking WebMCP supportChecking whether browser tools can register.
5agent tools
0approval tools
0dealer sends
Nothing has been sent to the dealership. This is a non-binding, in-memory challenge demonstration.
Try with an agent
Example prompts
- “Find performance sedans under €100,000.”
- “Compare KS-001 and KS-002.”
- “Show the condition of KS-001.”
- “Prepare €90,000, then evaluate only after I approve.”
Judge-ready manual fallback
Run the exact same functions on page
Use these controls if WebMCP is unavailable—or to inspect each trust boundary deliberately. The activity source stays honestly labeled manual.
- 1
- 2
- 3
- 4
- 5Try once before approval to prove the gate blocks it.
- 6Page-native human action. Never registered as a tool.
- 7
- 8
Shared visible state
Vehicle and decision workspace
Every tool or manual action updates this same page. Policy source, approval state, safe inputs, outcomes, and receipts stay inspectable.
Trust boundary
The agent can prepare. It cannot self-authorize.
Approval is intentionally outside the agent-callable tool surface. Even an override instruction cannot change the deterministic public policy. The receipt records what actually happened.