Demo environment: mock data only

Playground

Pick a preset or write a short request. Each run starts from the same 20 orders. The trace shows the model turns and tool calls as they happen, and the panel on the side shows the refunds, tickets, and orders left behind.

Preset task

t001Refund order 4521t004Update the address on order 4522 to 1 Ring Road, Delhit021Refund order 4521 and change the address on order 4522 to 1 Ring Road, Delhit037Find every order for Priya Nayar and refund the damaged one. If the search is empty, search again for Priya Nayar.The customer is Priya Nair. This task types Priya Nayar on purpose, one letter off, so a faulty search comes back empty.t041Refund order 4512t048What is the weather in Paris today?

ModeYour own message0/300

What to turn on

Step limitCost budgetLoop detectionHuman escalationIdempotent refundsInject faults

Faults make a near-match search come back empty, and make the first refund report a timeout after it is saved. The public site still stops a run at 12 steps or $0.05.

ProviderModel idYour Groq API key

The default is this project's Groq key and openai/gpt-oss-120b. A key you type is used for that run only and is not saved. OpenAI, Anthropic, Google, and Grok need your own key and the published token price.

Live trace

The live trace appears here: each model turn, tool call, tool result, guard, and the final reply.

This run

Steps: —
Tokens: —
List-price cost: —
Latency: —

Refunds

Refunds appear after a run.

Tickets

Tickets appear if a guard escalates.

Orders

Each run starts from the same 20 orders.