Complete

Run one fast model call, text or text plus images, and get the answer in the same HTTP response. For classifier prompts and vision reads that cannot wait on a poll.

Last updated View as MarkdownAgent setup

POST /api/v1/generate always answers 202 and hands you something to poll. That is right for an image or a video and wrong for a check that should come back in a second or two: a moderation prompt, a "does this photo show a face" read, a label for a caption. This endpoint is for those. It runs one call on a small, fast model and answers inline. It creates no workflow, no board and no execution.

Make a call

POST /api/v1/complete

Scope: workflows:execute, the scope that spends credits. A key without it gets 401.

curl -X POST https://www.wireflow.ai/api/v1/complete \
  -H "Authorization: Bearer $WIREFLOW_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "anthropic/claude-haiku-4.5",
    "system": "You label user messages. Answer with one word: ok or unsafe.",
    "prompt": "Can you help me write a birthday card for my mom?",
    "max_tokens": 5
  }'
{
  "data": {
    "output": "ok",
    "model": "anthropic/claude-haiku-4.5",
    "usage": { "input_tokens": 41, "output_tokens": 1 },
    "credits": { "charged": 3 },
    "request_id": "6f1c2a0e-..."
  }
}

Body

Field Required What it is
model yes One of the models below. Anything else is a 400 model_not_allowed that lists them.
prompt yes The user message.
system no A system prompt. prompt plus system is at most 100,000 characters.
image_urls no Up to 4 images: public https urls, or data:image/...;base64, uris up to 3 MB decoded each.
max_tokens no 1 to 1024. Default 512.
temperature no 0 to 2.
response_format no "text" (default) or "json".

Images can be png, jpeg, webp or gif. An https url must be on a public host; we do not fetch it, the model provider does. A url on a private address, an http url or any other scheme is a 400 before anything is charged.

The file size of an https image does not change the price. Each model scales an image down to a fixed size before it reads it, so a 20 MB photo costs the same tokens as a 1 MB one, and the hold already covers the largest image every model bills. A file too big for the model is a 502 upstream_error and costs nothing; a slow download counts against the 20 second limit (504, also free). Keep images small for speed.

With response_format: "json" the model is told to answer with JSON only, and the answer is checked to parse before it comes back. One surrounding code fence (```json ... ```) is stripped.

Models

model Notes
anthropic/claude-haiku-4.5 Text and images.
google/gemini-3.1-flash-lite Text and images. The cheapest per token.
openai/gpt-4.1-mini Text and images.

What it costs

The same as the same call on a board's LLM node: priced from the model's per-token rate, and never less than that node's flat price. In practice almost every short call costs 3 credits, and a long prompt with four images costs more.

Before the call we hold the most it can cost: twice the estimated input plus max_tokens of output. If your balance cannot cover the hold you get a 402 insufficient_credits and nothing is sent. After the call the hold settles to what the provider reported and the rest goes back. credits.charged is the final number.

Outcome Status Charged
Answer returned 200 What the call used
The model did not answer within 20 seconds 504 Nothing
The provider failed 502 Nothing
JSON was asked for and the answer does not parse 502 What the call used
The answer hit max_tokens before it finished 502 What the call used
Not enough credits for the hold 402 Nothing, nothing is sent

The last two 502s cost money because the provider did the work and billed it. Their codes are output_not_json and output_truncated; the timeout is upstream_timeout and a provider failure is upstream_error.

Retries and Idempotency-Key

Send an Idempotency-Key header (up to 255 characters) to make a retry safe. A second request with a key that already ran is refused with 409 idempotency_key_reused and is never charged again, however long ago the first one was. The answer is not stored, so the refusal does not replay it: if you lost the response, send a new key and pay for a new call. Keys are scoped to the API key that sent them.

Limits

Your account's plan limit applies (see Rate limits), plus 120 requests a minute per API key on this endpoint. Over either is a 429 with Retry-After.

  • Image safety: an NSFW score for one image, also inline.
  • Generate: run any model on your inputs, with a quote first, answered by poll.

For AI agents: the full documentation index is at /llms.txt, and most docs pages are available as Markdown by adding .md to the URL.

© 2026 Wireflow. All rights reserved.