I ve been building FetchSandbox, a sandbox layer that lets AI agents run real API workflows (Stripe, Clerk, Resend, etc.) and get back a verifiable receipt before touching production. Think Stripe Sandbox, but cross-vendor and designed for agents to drive over MCP.This playground has two tasks against a Descope integration. One asks your agent to add login and prove it works before writing any code. The other has a planted bug. Point Cursor or Claude Code at it, go in cold, and see what happens.Steps are in TESTING.md: https://github.com/fetchsandbox/playgroundNot looking for polish feedback. Specifically want to know: did the agent connect, did it catch the planted bug, where did it get confused or stall. The MCP surface is new enough that I expect rough edges and want to find them.