generated: '2026-08-27'
method: probed
source: https://pexafy.com/pricing/
probe: >-
  Response headers observed live on GET https://api.pexafy.com/health and on an
  unauthenticated GET https://api.pexafy.com/api/v1/search/photos, 2026-08-27.
limit_count: 7
headers:
  returned:
  - name: x-ratelimit-limit
    description: Requests allowed in the current window.
    observed_values: [100, 60]
  - name: x-ratelimit-remaining
    description: Requests left in the current window.
    observed_values: [98, 59]
  - name: x-ratelimit-reset
    description: Unix epoch seconds at which the window resets.
    observed_values: [1787852771]
  - name: x-request-id
    description: Correlation id, mirrored into meta.request_id in the body.
  retry_after: false
  retry_after_note: >-
    No Retry-After header was observed on the probes (neither returned a 429). The official
    Python SDK's changelog says "Retry-After honoured", so the server is expected to send it
    on an actual 429 — unconfirmed from an anonymous probe.
  standard: >-
    Lowercase X-RateLimit-* (the de-facto convention), NOT the IETF draft RateLimit-* /
    RateLimit-Policy form.
  status_on_exhaustion: 429
  error_code_on_exhaustion: RATE_LIMITED
  documented_in_openapi: >-
    The versioned description (openapi/pexafy-api-v1-openapi.json) documents a 429 response
    on every operation; the live served schema documents none.
observed:
- surface: GET https://api.pexafy.com/health
  scope: per-IP (unauthenticated)
  limit: 100
  window: unknown — x-ratelimit-reset was ~60s ahead at probe time
- surface: GET https://api.pexafy.com/api/v1/ (unauthenticated, 401)
  scope: per-IP (unauthenticated)
  limit: 60
  window: unknown — x-ratelimit-reset was ~60s ahead at probe time
  note: >-
    The unauthenticated API route is metered more tightly than /health. Both counters ran
    independently, so the limiter is per-route-class, not a single global bucket.
- surface: https://pexafy.com/ (website root, browser UA)
  status: 429
  note: >-
    The marketing site itself rate-limits aggressively; a HEAD from a scripted client
    returned 429. This is a site edge policy, not the API.
limits:
- scope: per-account (plan)
  plan: Free
  limit: 1200
  window: hour
  monthly_quota: 5000
  api_keys: 1
- scope: per-account (plan)
  plan: Starter
  limit: 1800
  window: hour
  monthly_quota: 20000
  api_keys: 3
- scope: per-account (plan)
  plan: Pro
  limit: 3600
  window: hour
  monthly_quota: 70000
  api_keys: 5
- scope: per-account (plan)
  plan: Expert
  limit: 100
  window: minute
  monthly_quota: 150000
  api_keys: 10
- scope: per-account (plan)
  plan: Team
  limit: 200
  window: minute
  monthly_quota: 300000
  api_keys: 25
- scope: per-account (plan)
  plan: Business
  limit: 300
  window: minute
  monthly_quota: 1000000
  api_keys: unlimited
- scope: per-account (plan)
  plan: Enterprise
  limit: 500
  window: minute
  monthly_quota: 2000000
  api_keys: unlimited
connector_limits:
  note: >-
    The pricing table lists a separate "Connector rate limit" for the WordPress / Chrome /
    Figma / MCP connectors, at the same numbers as the API rate limit on every tier
    (1,200/h Free, 1,800/h Starter, 3,600/h Pro, 100/min Expert, 200/min Team, 300/min
    Business, 500/min Enterprise). The UI itself is metered separately at 60/min on every tier.
self_service_quota_endpoints:
- GET /api/v1/usage
- GET /api/v1/usage/daily
- GET /api/v1/usage/monthly
- GET /api/v1/usage/by-key
quota_note: >-
  An agent can read its own consumption before it is throttled — the api-onboarding
  descriptor makes GET /api/v1/usage the third step of its first-run flow explicitly "so an
  agent can pace itself". The MCP server exposes no equivalent tool.
gaps:
- The docs page titled "Errors & rate limits" documents the 429 and the RATE_LIMITED code but publishes no numbers and does not name the X-RateLimit-* headers; the tiers are only on the pricing page and the headers only on the wire.
- The window length behind x-ratelimit-limit is not stated anywhere; it was inferred from the reset timestamp on a single probe and is recorded as unknown rather than guessed.