specification: API Commons Rate Limits
specificationVersion: '0.1'
schema: https://raw.githubusercontent.com/api-evangelist/interface-research/main/schema/api-commons.yml#/$defs/RateLimits
generated: '2026-08-30'
method: searched
source: https://agentgateway.dev/docs/standalone/latest/configuration/resiliency/rate-limits/ + .../operations/debug/
provider: AgentGateway
providerId: agentgateway
created: '2026-05-04'
modified: '2026-08-30'
description: >-
  Agentgateway imposes NO rate limits on callers, because there is no hosted agentgateway service to
  call - the product is a binary you run, and its own admin API is unauthenticated and unthrottled on
  loopback. Agentgateway is instead a rate LIMITER: the capability section below records the limiting
  it applies to traffic passing through it, which is what an integrator is actually asking about.
  THIS FILE REPLACES A SCAFFOLD. The previous version, from a bulk sweep dated 2026-05-04, asserted
  free/professional/enterprise tiers at 10/100/1000 requests per minute against an api-key scope, and
  X-RateLimit-* / RateLimit-Policy response headers. None of that was published by agentgateway. It
  has been removed.
limit_count: 0
limits: []
headers: {}
headers_note: >-
  No RateLimit-*, X-RateLimit-* or Retry-After response-header contract is documented for
  agentgateway's own limiter. This is a genuine gap in the product, and the single most useful runtime
  signal an agent is missing when it gets a 429 from an agentgateway-fronted route. Recorded as absent
  rather than assumed.
responseCodes:
  throttled: 429
  limiterUnavailable: 500
  limiterUnavailableNote: >-
    With the default failureMode failClosed, an unavailable or erroring remote rate limit service
    causes agentgateway to DENY the request with 500 Internal Server Error rather than 429 - so a 500
    from a rate-limited route is a limiter outage, not a gateway bug. failureMode failOpen allows
    traffic through instead.
capability:
  note: Rate limiting agentgateway APPLIES to proxied traffic. Configuration, not a limit on you.
  types:
    - id: local
      name: Local rate limits
      storage: in-memory, per replica
      shared: false
      detail: >-
        Very low overhead. Counters are not shared between replicas nor across restarts, so the docs
        state plainly they are not appropriate where exact global counts are required, or for long
        windows such as monthly limits.
      config_fields: [maxTokens, tokensPerFill, fillInterval, type]
    - id: remote
      name: Remote rate limits
      storage: pluggable external data store
      shared: true
      protocol: Envoy Rate Limit Service v3 gRPC
      protocol_url: https://www.envoyproxy.io/docs/envoy/latest/api-v3/service/ratelimit/v3/rls.proto
      detail: >-
        Chosen deliberately so existing Envoy rate limiters can be reused; the Envoy project's
        reference rate limiter with Redis is the documented example.
      config_fields: [host, domain, failureMode, descriptors, type]
      example: https://github.com/agentgateway/agentgateway/tree/main/examples/traffic-ratelimiting-global
  dimensions:
    - id: requests
      name: Request-based
      default: true
      detail: Each request consumes 1 unit of capacity.
    - id: tokens
      name: Token-based
      detail: >-
        Each prompt or completion token consumes 1 unit. Evaluated in two phases. At request time,
        with `tokenize: false` (the default) the token count is unknown so the request is always
        allowed unless the limit is 0; with `tokenize: true` agentgateway estimates and rejects if
        the estimate exceeds the limit. At response time the provider's reported counts are reconciled
        against the limit - which means a response that blows the budget is still returned, and only
        SUBSEQUENT requests are limited.
      caution: >-
        Since 1.5 the input count charged against a token limit INCLUDES cache-read and
        cache-creation tokens, normalised across providers. A limit sized against Anthropic or
        Amazon Bedrock (which exclude cached tokens from their own reported input count) now fills
        sooner than it used to.
  descriptors:
    detail: >-
      Remote limits key on CEL-valued descriptor entries, so limits can be scoped per organization,
      per authenticated user, per API key or per any header - e.g.
      value 'request.headers["x-organization"]'.
  attaches_to: [route]
  related_controls:
    - name: LLM budget and spend limits
      detail: Per-key dollar or token budgets and per-key model access lists, introduced in 1.5.
      docs: https://agentgateway.dev/docs/standalone/latest/llm/cost-controls/
    - name: Retries, timeouts, fault injection
      docs: https://agentgateway.dev/docs/standalone/latest/configuration/resiliency/
maintainers:
  - FN: Kin Lane
    email: kin@apievangelist.com