generated: '2026-09-02'
method: searched
source: https://standardcompute.com/getting-started
also:
  - https://standardcompute.com/smart-pacing
  - https://standardcompute.com/security
  - openapi/standard-compute-openapi.json
  - probed: POST https://api.stdcmpt.com/v1/chat/completions (2026-09-02)
summary: >-
  Standard Compute has almost no conventions of its own by design: it is a
  wire-compatible gateway, so the cross-cutting semantics an agent needs are
  OpenAI's (and Anthropic's), not this vendor's. The two places it does diverge are
  the model parameter (always the literal string "standardcompute" — the router
  chooses) and the exhaustion signal (HTTP 402 on monthly budget, never a 429).
auth:
  style: bearer-api-key
  header: 'Authorization: Bearer sc_live_...'
  see: authentication/standard-compute-authentication.yml
idempotency:
  supported: false
  header: null
  scope: null
  retention: null
  note: >-
    No idempotency key is documented, none appears in the OpenAPI, and none is
    referenced anywhere in llms.txt or llms-full.txt. Because completion endpoints
    are non-persisting the practical blast radius of a duplicate request is budget
    consumption rather than a duplicated record — real money, but not a duplicated
    business object. NO Idempotency pointer is wired into apis.yml: the check asks
    whether the provider supports idempotent retries, and this one does not.
  grade: absent
pagination:
  style: none
  note: >-
    No collection endpoint paginates. GET /v1/models returns the complete pool in a
    single {"object":"list","data":[...]} envelope with no cursor, and the pool is
    six entries.
expansion:
  supported: false
sparse_fieldsets:
  supported: false
metadata:
  supported: passthrough
  note: >-
    OpenAI request fields are forwarded to the selected upstream. GET /v1/models
    exposes per-model metadata the caller can branch on — context_length,
    supported_parameters (tools, tool_choice, response_format, reasoning,
    reasoning_effort), anthropic_family_tier and is_family_default. That is a
    genuinely useful capability-discovery surface and it answers unauthenticated.
request_id_tracing:
  client_supplied: false
  server_returned: not documented
  observed_headers:
    - x-cloud-trace-context
    - cf-ray
  note: >-
    Observed on a live 401. Both are infrastructure identifiers (Google Cloud Run
    and Cloudflare), not a documented, supportable correlation ID. Support asks
    users to paste error text and payloads rather than quoting a request id, which
    confirms none is published.
versioning:
  style: path-prefix (/v1)
  see: lifecycle/standard-compute-lifecycle.yml
error_envelope:
  shape: '{"error": {"message": "...", "type": "..."}}'
  rfc9457: false
  see: errors/standard-compute-problem-types.yml
rate_limit_signaling:
  headers: none
  status_on_exhaustion: 402
  status_429_used: false
  see: rate-limits/standard-compute-rate-limits.yml
  note: >-
    The single most consequential convention gap for an agent: budget state exists
    but is only rendered to a human in the dashboard. Nothing in the response tells
    a running agent it is close to a hard 402 stop.
content_negotiation:
  request: application/json
  streaming: >-
    Server-sent events, inherited from the OpenAI and Anthropic wire formats. Not
    described in the published OpenAPI, whose request and response bodies are empty
    schemas.
transport:
  tls_minimum: '1.2'
  plaintext_http: refused
  see: security/standard-compute-domain-security.yml
model_parameter:
  convention: >-
    Set model to the literal "standardcompute" on the OpenAI-compatible paths. The
    router selects the underlying model per request; the caller does not pin one.
    Named ids from GET /v1/models exist for callers that want a family tier.
  anthropic_path_exception: >-
    Claude Code and other Anthropic-Messages clients set ANTHROPIC_BASE_URL to
    https://api.stdcmpt.com with NO /v1 suffix, because the client appends
    /v1/messages itself, and set no model at all.
reversibility:
  applicable: na
  reason: >-
    Read-and-generate only. Every operation is an inference call — POST
    /v1/chat/completions, /v1/completions, /v1/responses, /v1/messages,
    /v1/messages/count_tokens — plus one read (GET /v1/models) and two static file
    reads. None of them creates, mutates or deletes a durable resource on Standard
    Compute's side, so there is no state for a reversal operation to act on. The
    provider explicitly retains no prompt or response content.
  write_surfaces: []
  reversal_operations: []
  irreversible_side_effect:
    description: >-
      One genuinely irreversible consequence exists and it is financial, not
      resource-shaped: a completed request consumes monthly compute budget, and
      budget consumption cannot be refunded, voided or reversed through the API.
      There is no cancel, no void and no per-request refund.
    mitigations_published:
      - name: Optional smart pacing
        effect: spreads the remaining budget across the month so a runaway agent cannot drain it in a burst
        source: https://standardcompute.com/smart-pacing
      - name: 7-day fair refund
        effect: >-
          Subscription-level, not request-level — cancel within 7 days and pay only
          what was used.
        source: https://standardcompute.com/pricing
    window: not applicable — no reversal path exists to have a window
  dry_run_mode:
    supported: partial
    operation: POST /v1/messages/count_tokens
    operationId: anthropic_count_tokens_v1_messages_count_tokens_post
    note: >-
      Token counting is the closest thing to a rehearsal: it lets an agent size a
      request before spending on it. It is Anthropic-wire-compatible and there is no
      OpenAI-path equivalent.
  grade: na
  grade_basis: >-
    Recorded as na rather than 0 per the contract: an API with no write surface has
    nothing to reverse, and an honest na leaves the denominator instead of scoring
    a zero the provider could not have avoided.