This experimental PPE plugin (ratelimit/limitador, embedded in-memory Limitador) caps how often each AI agent can call a shared model through the gateway. It counts admitted requests per verified identity and returns a real 429 at the gateway, before the call ever reaches the model.

Three panes: the governing policy snippet (top-left), live gateway activity showing each request admitted (↗ model) or denied (✗ no dispatch) (top-right), and the agent driving the calls (bottom).

Six acts:

  1. Accounting: the limiter counts each admitted call toward the per-agent budget.
  2. Admission gate: once the budget is spent, further calls get 429 and never reach the model.
  3. Identity-based keying: a second agent has its own counter, served while the first is throttled.
  4. Selection: a route with no limiter is never throttled.
  5. Route-specific limit: a different route carries its own, lower budget.
  6. Trust basis: no verified identity, no service. A missing token, or one with no subject, is denied first.