Send invoices in. Get data out. No downloads.

Quick start

Store your key in an environment variable first (export CONVERTPDF_API_KEY=…), then send an invoice:

printf '{"template":"invoice","pdfBase64":"%s"}' "$(base64 < invoice.pdf | tr -d '\n')" > request.json

curl -s https://convertpdf.pro/api/v1-extract \
  -H "Authorization: Bearer $CONVERTPDF_API_KEY" \
  -H "Content-Type: application/json" \
  -d @request.json

You get back one JSON object. Values are under “data”:

{
  "ok": true,
  "document": { "name": "document.pdf", "pageCount": 1 },
  "extracted_at": "2026-10-09T21:15:00.000Z",
  "model": "claude-sonnet-5",
  "data": {
    "vendor": "Northline Supply",
    "invoice_number": "INV-20458",
    "invoice_date": "2026-03-14",
    "total": 156.5,
    "line_items": [
      { "description": "Widget", "quantity": 4, "unit_price": 30, "amount": 120 },
      { "description": "Fastener kit", "quantity": 2, "unit_price": 18.25, "amount": 36.5 }
    ]
  },
  "details": { "vendor": { "label": "Vendor", "value": "Northline Supply", "status": "extracted" }, "...": "..." },
  "needs_review": false,
  "review_fields": [],
  "remaining_pages": 98,
  "charged": true,
  "read_method": "text_layer",
  "webhook": { "delivered": true, "attempts": 1, "status": 200, "error": null }
}

needs_review turns true when a value could not be confirmed or a required value (vendor, invoice number, total) is empty. review_fields says which ones, so your automation can route only those to a person.

Connect it to your tools

We have not published native Zapier or Make apps yet. Use their generic web-request step instead:

• Make: the HTTP module (“Make a request”). Method POST, header Authorization: Bearer your key, body type raw JSON.
• Zapier: Webhooks by Zapier, “Custom Request”, method POST (availability depends on your Zapier plan).
• n8n: the HTTP Request node, with the Extract From File node (operation “Move File to Base64 String”) in front of it.
• Your own code: any language that can send an HTTPS POST.

The PDF has to be sent as base64 text in pdfBase64, so add a step that converts the file to base64 first. For a photo or scan saved as an image, send it the same way in imageBase64 instead (send one or the other, not both).

Webhooks: push the result to your system

Add a webhookUrl and we also POST the result to it, signed so you can verify it.

{
  "template": "invoice",
  "pdfBase64": "<the PDF as base64 text>",
  "webhookUrl": "https://hook.example.com/your-endpoint"
}

The webhook receives the same JSON as the response above (without the "webhook" field), plus "event": "extraction.completed".

• The URL must be https on a public hostname (no IP addresses, no internal or local addresses).
• Headers sent: X-ConvertPDF-Signature, X-ConvertPDF-Timestamp, X-ConvertPDF-Delivery-Id (the same id on a retry, so you can ignore duplicates).
• If your endpoint answers with a server error (5xx), “too many requests” (429), a network error or does not answer within 6 seconds, we try once more after a second. Other error codes are not retried. The outcome is shown in the webhook field of the response, and a failed delivery never costs you extra pages.
• Most digital PDFs finish within a few seconds to half a minute. Scans take longer because they are read first, and in the worst case one request can run for close to four minutes. Set your tool’s timeout as high as it allows (at least 240 seconds). Your hosting limits may cut a request short before that, so test with a real scan before relying on it. A webhook is sent from inside the same request, so it does not help if your tool gives up early, and there is no background mode yet. If your tool cannot wait that long, scans may not suit it yet. A request that your tool abandoned may still be completed and charged.
• Checking the signature is optional but recommended for anything that writes to your books. It needs the raw request body, which some no-code “catch hook” steps do not expose; the example below is for your own code.

import { createHmac, timingSafeEqual } from "node:crypto";

// rawBody must be the exact request body text, before any JSON parsing.
function isValid(rawBody, headers, webhookSecret) {
  const timestamp = Number(headers["x-convertpdf-timestamp"]);
  const signature = headers["x-convertpdf-signature"] ?? "";
  if (!Number.isFinite(timestamp) || Math.abs(Date.now() / 1000 - timestamp) > 300) return false; // missing or too old
  const expected =
    "sha256=" + createHmac("sha256", webhookSecret).update(timestamp + "." + rawBody).digest("hex");
  return signature.length === expected.length &&
    timingSafeEqual(Buffer.from(signature), Buffer.from(expected));
}

Errors

Limits

• Scanned PDFs and photos (JPEG, PNG, WebP) work too, up to 10 pages, and are read by AI before extraction. The result includes "read_method": "ai_vision". Scans are less exact than digital PDFs, so rely on the needs_review flag. A PDF that has any scanned page is read entirely as a scan (10-page limit), and the file itself is sent to the AI provider.
• Up to 20 pages, about 3 MB and about 50,000 characters of text per PDF.
• Paid page packs only; pages are deducted exactly like the web app.
• If nothing at all is found in a document (every field empty), it is not charged and the response shows "charged": false. This applies to up to 5 such results per day per account; after that they are charged like any other document.
• The first time the same document text is sent again, it is a free redo instead of being charged again. It is applied automatically, so an automatic retry in your tool can use it up; later repeats are charged. Scans are read by AI each time, so their text is rarely identical and they rarely qualify.
• Send the PDF itself — we do not fetch PDFs from links.

Privacy: unlike the in-browser converter, the API receives your PDF on our server so it can read the text. The extracted text is then sent securely to an AI provider to find the fields. We do not intentionally store that document text afterward, and we do not guarantee how long the AI provider itself retains it. For scans and photos, the file itself is sent to the AI provider so it can be read.