Skip to main content
Hue simulates the apps your agent works with, so you can rerun a real user’s task against apps that behave like production without touching anyone’s real accounts. A simulation answers your agent’s API and MCP calls in each app’s own format, from an isolated copy of the app’s data, and records every call and change.

​
How simulations work

  • The app’s own interface. A simulated app serves supported API paths and MCP tools in the app’s request and response formats. Your agent keeps its existing client; only the base URL and credential change during an evaluation.
  • State that carries over. Calls read and change a World, the starting state of the user’s apps, so a later call sees what an earlier one changed, as it would in production.
  • Isolated attempts. Each case of a run gets its own World instance, so attempts never affect each other or a real account, and each one stays inspectable afterwards.
  • Evidence for grading. Hue records every call and the resulting state, and evaluators grade what the agent actually did rather than what it said it did.
Cases with a World explains how a case handles a call to a tool Hue doesn’t simulate.

​
Apps

Call list_apps over Hue MCP for the current catalog, then get_app for an app’s API and MCP connections. get_world_apps lists the operations a specific World version supports and its known gaps. Organizations with Simulations access can also inspect each app’s fidelity results in Hue.

​
Connect your agent

Connect your agent adds one helper that gives your agent’s app clients Hue’s URLs and the world token during an evaluation, and its production values otherwise. To try a simulated app by hand, start a World instance and use its connection values, as Worlds describes. Worlds explains how Hue builds a World from your traces and what to record for it. The TypeScript runSimulation runner, World definitions, the callback context, the World handoff and their limits are in the TypeScript API reference.