1. Base URL & Authentication
Connect official OpenAI SDKs, evaluation harnesses, or HTTP clients using a workspace virtual key.
Environment API Base URL:
https://dotprod.tech/v1
Authentication Header:
Authorization: Bearer sk-...
Supported endpoints: GET /v1/models, POST /v1/sessions, POST /v1/chat/completions, POST /v1/responses (for Codex CLI and OpenAI Responses clients), and POST /v1/messages / POST /v1/messages/count_tokens (for Claude Code and Anthropic Messages clients). Unexposed routes (such as /v1/embeddings or /v1/completions) return HTTP 404.
Supported Models on This Gateway
2. Chat Completions & Buffered SSE Screening Behavior
POST /v1/chat/completions — Plain-text chat messages with pre-inference ceiling holds and safety screening.
- Buffered SSE delivery (full-output pre-release screening): When
stream: trueis passed on/v1/chat/completions,/v1/responses, or/v1/messages, DotProd returns standard wire-compatibletext/event-streamframes so streaming SDKs and coding harnesses (Claude Code, Codex CLI, OpenAI SDK) work without code changes, after buffering, R2-capturing, and screening the complete model response across overlapping moderation windows. First-chunk latency therefore equals full generation + screening time. - Text, tool calls & structured outputs: Plain-text messages, function tool calls (
tools/tool_choice), and structured outputs (response_format:text,json_object,json_schema) are screened, captured, and budgeted; multimodal binary attachments and upstream persistence (store: true) are rejected with HTTP 400. - Usage reporting: Uncached vs. prefix-cached input splits are reported when provided by the upstream model server; when omitted, DotProd records
null(displayed asunknown) rather than reporting0, and reconciles missing usage provisionally until authoritative provider logs arrive.
Python (OpenAI SDK)
cURL Example
3. Spend Controls & Failure Responses
Prepaid credit balances and UTC monthly spend budgets are enforced independently inside TenantLedgerDO.
- 400 Bad Request — Invalid UTF-8 JSON,
max_tokensout of range, unsupported non-text/tool parameters, or input/output blocked by content policy (code: "content_blocked"). - 401 Unauthorized — Missing, invalid, blocked, or revoked
sk-...virtual API key (type: "authentication_error"). - 402 Payment Required — Prepaid credit balance is below the conservative request reservation ceiling (
code: "insufficient_funds"). - 403 Forbidden — Requested model is not enabled for your workspace, or the workspace is suspended.
- 429 Too Many Requests — Aggregate UTC monthly spend budget reached (
code: "budget_exceeded") or per-minute RPM limit exceeded (code: "rate_limit_exceeded", includesRetry-Afterheader). - 502 / 503 — Upstream inference failure or unavailable content-retention/screening dependency (fail-closed).
Example Error Payload (HTTP 429 Monthly Budget Cap)
{
"error": {
"message": "Monthly spend budget (2026-11 UTC) of $50.00 reached ($49.9200 settled/held)",
"type": "insufficient_quota",
"code": "budget_exceeded"
}
}
4. Content Capture, Privacy & Retention Terms
How request inputs, model outputs, and waitlist submissions are stored and audited.
Inference content retention: Before model dispatch and before response release, DotProd captures plain-text request inputs and model outputs to Cloudflare R2 encrypted with AES-256-GCM per-object envelope keys, paired with SHA-256 integrity digests in D1. Raw captures are never exposed via shared admin keys; decryption requires a named, MFA-verified reviewer identity holding an explicit can_read_content grant for the target tenant and a mandatory 8–500 character incident/ticket reason recorded in access_audit.
Waitlist submissions: Email addresses and plan preferences submitted to /waitlist are protected by Cloudflare Turnstile, rate-limited using salted hashes, retained until their configured expiry window (purged automatically by scheduled Worker tasks), and accessible only to MFA-verified staff with audited role grants.