Developer API

API reference

Overview

EdifacePay exposes a documented subset of the OpenAI Chat Completions request shape: listing models and creating text chat completions, with or without streaming. Requests are billed to your prepaid balance through the same reservation and ledger path as the workspace.

It is not a full OpenAI-compatible API, and it is not compatible with the Anthropic SDK or with tools that require features outside this page. Parameters that are not listed here are rejected with HTTP 400 rather than silently ignored, so a request that succeeds did what it said.

Base URL: https://www.edifacepay.org/v1

Authentication

Create a key under API keys. The secret is shown once; only a keyed hash is stored. Send it as a Bearer token:

Authorization: Bearer efp_…

Revoked, expired or unknown keys receive 401. Keys of a suspended account receive 403. Provider credentials are never exposed — your key only authenticates you to EdifacePay.

GET /v1/models

Returns the models that can be called right now (enabled, provider configured, last live test passed).

curl https://www.edifacepay.org/v1/models -H "Authorization: Bearer $EDIFACEPAY_API_KEY"

{ "object": "list",
  "data": [{ "id": "claude-opus-5", "object": "model", "owned_by": "…",
             "context_window": 200000, "max_output_tokens": 64000 }] }

POST /v1/chat/completions

curl https://www.edifacepay.org/v1/chat/completions \
  -H "Authorization: Bearer $EDIFACEPAY_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: 6f1c…"   # optional \
  -d '{ "model": "claude-opus-5",
        "max_completion_tokens": 1024,
        "messages": [
          { "role": "system", "content": "Be concise." },
          { "role": "user", "content": "Explain idempotency in one paragraph." } ] }'

Response (non-streaming):

{ "id": "chatcmpl-req_…", "object": "chat.completion", "model": "claude-opus-5",
  "choices": [{ "index": 0, "finish_reason": "stop",
                "message": { "role": "assistant", "content": "…" } }],
  "usage": { "prompt_tokens": 31, "completion_tokens": 58, "total_tokens": 89,
             "prompt_tokens_details": { "cached_tokens": 0 } },
  "edifacepay": { "request_id": "req_…", "status": "settled", "cost_usd": "0.000321" } }

finish_reason is stop, length (output limit reached) or content_filter (the model declined). The edifacepay object is an extension: the request id you will find under Usage, and the exact amount debited. The header x-edifacepay-request-id carries the same id.

Streaming

With "stream": true the response is text/event-stream of chat.completion.chunk objects, ending with data: [DONE]. With include_usage, the last chunk before [DONE] has empty choices and carries usage and edifacepay. If the provider fails mid-stream, the stream ends with a single {"error":…} object and no [DONE].

Disconnecting cancels. If you close the connection, the upstream generation is aborted. Providers bill what was generated before the abort and do not report it; see below.

Billing behaviour

• Reserve: before the provider is called, a conservative maximum (estimated input + the full output allowance) is reserved atomically. If your available balance, your account limits or the key’s limits cannot cover it, the request is refused with 402 and nothing is sent. Lower max_completion_tokens to reserve less.
• Settle: cost = provider-reported tokens × the price version captured at reservation, summed over input, cached input, cache write and output, rounded up once to $0.000001. The unused part of the reservation is released immediately.
• Uncertain usage: if a stream is cancelled or interrupted before the provider reports usage, the reservation stays held as pending_reconciliation, then settles from an estimate (≈4 bytes per token over what was actually received, capped at the reservation) that is marked estimated on the usage record. Requests the provider rejected outright are not charged.
• No automatic retries. EdifacePay never re-sends a generation that may already have been billed upstream. Retry from your side; use Idempotency-Key so an accidental duplicate is refused with 409 instead of being billed twice.

Errors & limits

{ "error": { "message": "…", "type": "invalid_request_error", "param": "tools", "code": "unsupported_parameter" } }

A single generation is stopped after 280 seconds.

Not supported

Tool and function calling, JSON mode and structured outputs, images, audio and files, stop, seed, logprobs, logit_bias, penalties, reasoning_effort, n > 1, stored completions, embeddings, the Responses API, the Assistants API, batch and fine-tuning endpoints. Each is rejected with 400.