The AI detector API that shows its evidence.

Import the OpenAPI contract

The OpenAPI 3.1 JSON documents POST /v1/check, POST /v1/prompt, POST /api/v1/c2pa/inspect and GET /health, with a reference entry for the POST /mcp endpoint, including request formats, check results, image diagnostics, saved-prompt responses and every error, down to a missing saved record. Import its URL into an API client or feed it to a compatible client generator. The other endpoints below sit outside this contract.

No API key is required for these documented endpoints. For a private /v1/check request, send privacy=private and public_receipt=false; public URLs default to public checks, so set this deliberately. For saved-prompt extraction with no model call, send mode=saved_only to /v1/prompt. Private uploads are processed on isGenAI’s server, not in your browser, and stay out of public reports.

curl --fail-with-body --max-time 30 https://isgenai.com/v1/check \
  -F "file=@picture.png" \
  -F "privacy=private" -F "public_receipt=false"

JSON bodies go up to 128 KiB. Files go up to 20 MiB, with 64 KiB of additional multipart overhead. Text and code go up to 16,000 UTF-16 code units: /v1/check trims surrounding whitespace before measuring length, while /v1/prompt counts the original input. HTTP 429 means shared capacity is busy or a reconstruction limit has been reached; respect Retry-After when present and skip automatic retry loops. Every successful /v1/check response lists which adapters ran, so an unavailable adapter or an open origin is reported as exactly that.

Handle failures with the stable error.code, such as missing_input, too_large or busy, together with the HTTP status. error.message is human-readable context, so branch on the code rather than the wording. Use a generic failure message for absent or unfamiliar codes, and never infer an image’s origin from an error.

Use isGenAI from Claude, ChatGPT and Cursor (MCP)

The same checks are available as a public remote MCP server at https://isgenai.com/mcp (Streamable HTTP, no account or API key). Five read-only tools check text, public links, images, saved prompts and C2PA Content Credentials, and return each finding with its source and the report link.

claude mcp add --transport http isgenai https://isgenai.com/mcp

Set up the MCP server in Claude, ChatGPT, Cursor or VS Code.

Inspect a TIFF without converting the original

Send a single-page classic TIFF as the multipart file to POST /v1/check. The checker preserves the original bytes for the file hash and saved-record checks, and computes image diagnostics and a visual fingerprint from the same file. Use a file up to 20 MiB and 40 megapixels; multipage TIFF, TIFF subimages/pyramids, BigTIFF and unreadable files return a clean HTTP 400.

TIFF gets the full file-evidence and diagnostics path; the statistical image classifier and prompt reconstruction are reported as unsupported for this format. A saved AI declaration is reported as editable evidence, and a file with no record is reported as exactly that. Follow the TIFF inspection guide and controlled example.

Read the prompt saved in an image

Send the original image to POST /v1/prompt as a multipart file. isGenAI reads AUTOMATIC1111 settings in PNG, JPEG and WebP files, ComfyUI workflows in PNG files including Flux, and InvokeAI creation settings in PNG files, word for word.

curl --fail-with-body --max-time 30 https://isgenai.com/v1/prompt \
  -F "file=@picture.png" -F "mode=saved_only"

A response with status: "saved" contains the stored prompt and, when the file carries them, negative_prompt and recorded_model. Reading saved details uses no AI model call. The file must still carry those details; a workflow with several prompts and no single output prompt is reported as exactly that.

Use mode=saved_only to extract saved details without ever sending the image to an AI model. When none are found, the endpoint returns HTTP 422 with error.code: "no_saved_prompt": a missing file record, cleanly separated from any AI or human verdict. Download the Node.js extraction example.

Omit mode or use auto to allow a new suggestion when no saved prompt is available. That response has status: "reconstructed" and is a prompt for a similar result, labelled as such. Reconstruction is subject to availability and service limits; handle non-success responses without automatic retry loops.

Images go up to 20 MiB (20,971,520 bytes). The saved-prompt guide covers original downloads, screenshots and supported workflows.

For a small set of files, follow the batch saved-prompt example. It checks up to 10 explicit files sequentially, keeps missing records separate from failures and writes a private JSON report without ever overwriting an existing file.

Image evidence and quality in one response

Image checks include image_diagnostics: versioned dimensions, resolution, alpha-weighted brightness, tonal spread, near-black and near-white pixel percentages, transparency and neighboring-pixel differences. Measurements run on isGenAI’s server with no separate model call.

Read status before accessing measurements. An unavailable analysis reports its reason; fully transparent samples return null for brightness and edge measurements. The version 1 sample fits within 512 × 512 pixels, corrects EXIF orientation and covers the first frame. Measurements describe the pixels and sit apart from the origin verdict.