generated: '2026-08-09'
method: searched
source:
  - https://docs.routerplex.com/authentication
  - https://docs.routerplex.com/errors
  - https://docs.routerplex.com/streaming
  - https://docs.routerplex.com/keys-budgets
  - openapi/routerplex-inference-openapi.json
authentication:
  style: static API key
  transports:
    - {header: Authorization, format: 'Bearer sk-...', clients: OpenAI-compatible}
    - {header: x-api-key, format: 'sk-...', clients: Anthropic-compatible}
  note: >-
    Both styles are accepted on /v1/messages. Keys are created and revoked in the
    dashboard; revocation is immediate. There is no OAuth, no token refresh, no
    expiry — the key is the whole credential.
  key_scoping:
    per_key_budget: true
    per_key_model_allowlist: true
    per_key_rpm_tpm: optional
    note: >-
      Documented at https://docs.routerplex.com/keys-budgets as the recommended
      pattern for agents: give each app, IDE or agent its own key with its own hard
      lifetime budget and model allowlist so "an agent gone wild can't drain your
      whole balance."
  see_also: authentication/routerplex-authentication.yml
idempotency:
  supported: false
  header: null
  note: >-
    No idempotency key is documented or present in the OpenAPI. Inference requests
    are non-idempotent and billable — a retried POST is a second charge. This is
    inherited from the OpenAI/Anthropic wire formats, which also have no idempotency
    contract on completions. Agents must dedupe client-side.
pagination:
  supported: false
  note: >-
    No paginated collections. GET /v1/models returns the full list for the key in
    one response.
field_expansion:
  supported: false
metadata:
  supported: passthrough
  note: >-
    Request schemas are additionalProperties:true throughout, so vendor-specific
    fields pass through to the upstream model provider. RouterPlex does not define
    its own metadata object.
request_tracing:
  request_id_header: null
  note: >-
    No RouterPlex request-id header is documented. Cloudflare's cf-ray is present on
    responses but is edge infrastructure, not a provider-supported correlation ID.
versioning:
  style: uri-path
  current: v1
  see_also: lifecycle/routerplex-lifecycle.yml
error_envelope:
  format: openai
  shape: '{"error":{"message","type","code"}}'
  content_type: application/json
  rfc9457: false
  see_also: errors/routerplex-problem-types.yml
rate_limit_signaling:
  headers_documented: false
  note: >-
    No RateLimit / X-RateLimit / Retry-After headers are documented. Limits surface
    only as a 429 with a message. Retry guidance is prose: retry a rate limit, do not
    retry a reached spend limit.
  see_also: rate-limits/routerplex-rate-limits.yml
streaming:
  style: server-sent-events
  trigger: 'stream: true'
  max_duration: 10 minutes
  note: >-
    The /messages route emits Anthropic SSE event names (message_start,
    content_block_delta, message_stop) for EVERY model, including non-Claude models —
    the gateway translates the event stream, not just the request shape.
compatibility_contract:
  note: >-
    RouterPlex's core convention is that it has no conventions of its own. It adopts
    OpenAI's and Anthropic's wire formats verbatim so the official SDKs work with only
    base_url changed. Portability is the product; the tradeoff is that it also
    inherits their gaps — no idempotency, no problem+json, no rate-limit headers.