generated: '2026-07-18'
method: searched
source: https://cumuluslabs.io/
docs: https://docs.cumuluslabs.io/
summary: >-
  Cross-cutting request/response semantics for the Cumulus inference gateway,
  captured from the public marketing site and docs index. The gateway is an
  OpenAI-compatible HTTP layer, so wire-level request/response shapes follow the
  OpenAI Chat Completions / Responses conventions; Cumulus adds declared
  per-workflow routing on top.
authentication:
  style: bearer
  header: Authorization
  format: Bearer <CUMULUS_KEY>
  ref: authentication/cumulus-labs-authentication.yml
compatibility:
  openai_compatible: true
  base_url: https://api.cumuluslabs.io/v1
  drop_in_for: [OpenAI SDK, Anthropic SDK, LangChain, LlamaIndex, Vercel AI SDK]
  note: >-
    "Change one line" — swap the OpenAI base_url to https://api.cumuluslabs.io/v1
    and keep existing client code. Request/response envelopes mirror the
    upstream provider being emulated.
routing:
  style: declared-per-workflow
  description: >-
    Routing rules are declared per workflow and pick the model, provider, and
    infrastructure per request; Cumulus describes this routing as deterministic
    and traceable, with every request logged (input, output, model, latency,
    cost, quality) to a replayable audit log.
caching:
  layers: [exact-match, prefix, semantic]
  note: Stacked prompt/KV cache; semantic cache optional.
versioning:
  scheme: uri-path
  current: v1
  evidence: base URL https://api.cumuluslabs.io/v1
idempotency:
  supported: unknown
  note: >-
    No idempotency-key contract is documented on the public surface reviewed;
    the accessible docs are behind a bot wall. Not asserted — do not treat as
    supported without confirmation.
observability:
  request_log: true
  fields: [input, output, model, latency, cost, quality]