generated: '2026-09-19'
method: probed
source: >-
  Response headers observed on live unauthenticated requests 2026-09-19, cross-checked against the
  one place the provider states a limit in words (agents.json rate_limits) and the A2A card's
  x_rs_ai_friendly block ("fair-use only").
docs: https://rsperformance.online/.well-known/agents.json
limit_count: 2
limits:
  - scope: per-client (keying not stated; Laravel throttle behaviour implies per IP)
    surface: A2A JSON-RPC — POST https://rsperformance.online/
    window: 1 minute
    limit: 60
    burst: null
    documented: 'agents.json rate_limits.requests_per_minute: 60 ("Soft fair-use limit on shared host")'
    observed_headers:
      X-RateLimit-Limit: '60'
      X-RateLimit-Remaining: '59'
    status_on_exhaustion: not observed
    retry_after: not observed
  - scope: per-client
    surface: REST apex — GET https://rsperformance.online/api/knowledge/search
    window: not stated (Laravel-style per-minute headers)
    limit: 30
    burst: null
    documented: no — header only
    observed_headers:
      X-RateLimit-Limit: '30'
      X-RateLimit-Remaining: '27'
    status_on_exhaustion: not observed
    retry_after: not observed
unmetered_or_unsignalled:
  - surface: REST gateway — POST https://ai.rsperformance.online/api/search
    observed_headers: none (no X-RateLimit-* or RateLimit-* on a 200)
    documented: 'no numeric limit; the provider steers RAG sweeps here ("Prefer VPS gateway ... for RAG sweeps", security.txt)'
  - surface: MCP — https://mcp.rs3d.pl/
    observed_headers: none
    documented: no
  - surface: static well-known documents and feeds
    documented: 'cache TTL guidance only — >= 600 s on /.well-known/* (security.txt, agent card); beacon Cache-Control public, max-age=300'
headers:
  family: X-RateLimit-* (de facto; not the IETF RateLimit-* draft)
  returned: [X-RateLimit-Limit, X-RateLimit-Remaining]
  not_returned: [X-RateLimit-Reset, RateLimit-Policy, Retry-After]
note: >-
  The runtime signal exists only on the canonical host. An agent calling the gateway or the MCP
  server gets no header telling it how close it is to a limit and must rely on the prose fair-use
  policy. Exhaustion behaviour (429 or otherwise) was deliberately not provoked.