Three pre-loaded scenarios — supplier-payment agent, AML investigation copilot, customer-support agent. Each ships with hundreds of observed actions, detected Authority Gaps, Guard Recommendations, and a signed Evidence Pack you can inspect. No login. No data leaves your browser.
An AP-automation agent at Meridian Mutual Bank. Reads invoices, matches POs, proposes payments. The fixture shows 12 Authority Gaps including a payment.execute attempt outside scope.
A SAR-drafting copilot at Halberd Sterling Bank. Critical: 3 case.close attempts blocked, 7 SAR drafts contained unredacted PII. Demonstrates the redaction + denied-action gates.
A clean-baseline fixture from Cobalt Energy Retail. Shows what healthy looks like — 5,000 actions, only 2 minor gaps, every conversation recorded with a Consent Receipt.
Construct an observed_action by hand, submit it, see the live shadow-mode evaluation. Backed by the same precedence rules a production gateway would apply — running entirely in your browser.
Connect from Claude Desktop, Cursor, or Cline. The sandbox exposes a live Model Context Protocol endpoint at mcp.sandbox.kyeprotocol.com with all 12 KYE tools mounted.
Every observed action is bound to a signed Evidence Pack. Open one to see the authority chain, scope check, drift score, and transparency-log sequence.
Take any fixture action, modify its context (raise the risk score, flip external_export to true), see the verdict change live.
Every artefact is shareable. Take a screenshot, link to the scenario, or follow Apply for pilot → to bring KYE to your real AI-agent traffic.
Skip the sandbox — go directly to the OSS path. The @kye/shadow-mode-sdk ships under Apache-2.0 and integrates into your AI-agent runtime in five minutes.