Guarantees
What a mock guarantees
A mock is only worth anything if your agent can't tell the difference. This page says what you can rely on for that, and what you can't.
The surface is the real one
A tool you see in Mockd carries the real provider's description and schema, unchanged down to the wording and the fields.
That's the whole point. The tool description is the text your model makes its decision on. A re-typed description would distort exactly the thing you came to test.
State: calls affect each other
The data is synthetic, but it isn't new on every call. What your agent creates, changes or links stays there for the rest of the session.
Without that, a mock is a vending machine. An agent that creates a contact and then can't find it behaves differently from one that can — and that behaviour is what you want to see before real data is involved.
Deterministic from the start
Every session starts from the same records with the same IDs — every time.
Latency isn't random either: the same call in the same session always takes the same time, normally around 90 milliseconds. Useful when you're comparing two runs.
What a mock doesn't do
- It doesn't invent tools. What the real server can't do, the mock can't either — not even when it would be convenient.
- It doesn't call a real API. There's no account behind it, no key and no quota.
- It isn't the real service. Mockd has no connection to the providers whose servers are mocked.
The catalog grows
A new server shows up in the catalog as soon as it is supported, with its real tool names. Which ones are mockable today and which are announced is listed there, served straight from the backend.