PDF Text Extractor + OCR for Scanned PDFs

Give it PDF links, get back the text. Pages that are scans or photos are read with OCR automatically; pages that already have text are never OCRed. One flat row per URL, every field always present, never null — built to be called from pipelines, n8n/Make/Zapier flows and AI agents.

$2 per 1,000 PDFs. $5 per 1,000 OCR pages. Failed downloads, broken files and PDFs with no text are free.

Why this one

• It does not fall over on scans. OCR runs one page at a time with bounded memory. On a 41-file batch with 120 scanned pages, peak memory was 541 MB at the default 2 GB.
• One row for every URL you send. A 404, a password-protected file and a corrupt PDF each still come back as a row with status: "failed" and a machine-readable subStatus. You never have to diff your input against the output to find what went missing.
• You pay for text, not for attempts. A PDF is billed only when text came out of it. An OCR page is billed only when OCR recovered more text than the page already had.
• Fast. The text layer is read with PDFium, the PDF engine inside Chrome: 1–8 ms per page. An 838-page book took 18 seconds end to end, including the 52 MB download.
• Share links just work. Google Drive, Dropbox, OneDrive and arXiv /abs/ links are turned into direct downloads. If a host blocks the request (401/403/407/429), the download is retried once through Apify Proxy.

Input

Field	Default	What it does
urls	—	PDF links, one per line. Required.
ocrMode	auto	auto OCRs only pages with no text layer that contain an image. off never OCRs.
ocrLanguage	eng	Tesseract language code; join several with + (eng+fra). 30 languages are built in, including vie, chi_sim, jpn, kor, ara, hin.
maxOcrPagesPerPdf	50	Cost cap for huge scanned books. Pages over the cap come back empty and are listed in notes.
pageRange	all	1-5, 8, 12- style. A range that selects no page fails that row, uncharged.
includePages	false	Adds pages[] with per-page text and an ocr flag, for page-level citations.
chunkSize / chunkOverlap	0 / 200	RAG chunks of about chunkSize characters, each with pageStart and pageEnd. 0 turns chunking off.
password	—	Used for every PDF in the run that needs one.
maxFileSizeMb	100	Larger files fail as too_large, uncharged.

{
  "urls": ["https://arxiv.org/abs/1706.03762", "https://www.irs.gov/pub/irs-pdf/f1040.pdf"],
  "ocrLanguage": "eng",
  "chunkSize": 1500
}

Output

One row per input URL, in completion order. Use inputIndex to line rows up with your input.

{
  "inputIndex": 0,
  "url": "https://arxiv.org/abs/1706.03762",
  "resolvedUrl": "https://arxiv.org/pdf/1706.03762.pdf",
  "fileName": "1706.03762.pdf",
  "status": "ok",
  "subStatus": "text_layer",
  "error": "",
  "pageCount": 15, "pagesExtracted": 15, "ocrPages": 0, "scannedPages": 0,
  "charCount": 39598, "wordCount": 6099,
  "text": "Attention Is All You Need\nAshish Vaswani∗\nGoogle Brain\n…",
  "textTruncated": false, "fullTextUrl": "",
  "pages": [], "chunks": [],
  "title": "", "author": "", "subject": "", "keywords": "", "creator": "LaTeX with hyperref", "producer": "pdfTeX-1.40.25",
  "creationDate": "2024-04-10T21:11:43Z", "modDate": "2024-04-10T21:11:43Z",
  "pdfVersion": 15, "isEncrypted": false, "fileSizeBytes": 2215244,
  "notes": [], "processingMs": 218
}

Strings are "" when absent, numbers 0, arrays []. No field is ever null or missing.

Status

status	subStatus	Charged
ok	text_layer — all text came from the PDF itself	PDF
ok	ocr — some or all pages were read by OCR	PDF + each OCR page
ok	ocr_incomplete — some scanned pages were not read (page cap, cost limit, OCR off, OCR failure); see notes	PDF + each OCR page
no_text	no_text_layer, scanned_ocr_off, ocr_found_nothing	free
failed	http_404 (any http_<code>), not_a_pdf, too_large, download_timeout, download_error, password_required, wrong_password, unsupported_encryption, corrupt_pdf, empty_document, bad_page_range, page_range_out_of_bounds, invalid_url, cost_limit_reached, extraction_error	free

Very large documents: if a row would exceed Apify's dataset item limit, the full text is saved to the run's key-value store and linked from fullTextUrl. text keeps the first 1,000,000 characters and textTruncated is true.

Pricing

Event	Price	When
pdf-extracted	$0.002	A PDF returned text (status: "ok")
ocr-page	$0.005	A scanned page whose text was recovered by OCR

The only other charge is Apify's standard Actor start fee, $0.00005 per GB of memory ($0.0001 per run at the default 2 GB). Prices drop on higher Apify plans: down to $1.50 per 1,000 PDFs and $3.50 per 1,000 OCR pages on Gold. A run stops starting new PDFs when your maximum cost per run is reached, and marks the rest cost_limit_reached, uncharged.

What it does not do (yet)

Honest limits, so you can decide before you run it:

• Plain text, not Markdown. Line breaks are kept and pages are joined with a blank line, but headings and lists are not marked up.
• Tables come out as text, one table row per line. Cells are not split into columns.
• Text order follows the PDF's own content order. For most papers, reports and forms that is reading order. Some designed layouts (magazines, brochures) can interleave columns.
• OCR is Tesseract with fast models. It is good on clean scans and weaker on handwriting, very low-resolution photos and heavily stylised fonts.
• Text-based PDFs only need small memory. Scans need CPU. The default 2 GB OCRs about one page every 4–5 seconds. For big scanned batches, 8 GB runs two OCR pages at a time.

Use it from AI agents (MCP)

This Actor is a tool on the Apify MCP server, so Claude, Cursor, VS Code or any MCP client can call it to read a PDF. Load only this tool: a short tool list keeps the agent from picking the wrong one.

Recommended: hosted server, OAuth sign-in (nothing secret on disk). Add to your MCP client config (claude_desktop_config.json, .cursor/mcp.json, …):

{
  "mcpServers": {
    "pdf-text-extractor": {
      "url": "https://mcp.apify.com?tools=sumitr_mardy/pdf-text-extractor"
    }
  }
}

With an API token instead (clients without OAuth, or unattended setups):

{
  "mcpServers": {
    "pdf-text-extractor": {
      "url": "https://mcp.apify.com?tools=sumitr_mardy/pdf-text-extractor",
      "headers": { "Authorization": "Bearer YOUR_APIFY_TOKEN" }
    }
  }
}

Local server (stdio):

{
  "mcpServers": {
    "pdf-text-extractor": {
      "command": "npx",
      "args": ["-y", "@apify/actors-mcp-server@latest", "--tools", "sumitr_mardy/pdf-text-extractor"],
      "env": { "APIFY_TOKEN": "YOUR_APIFY_TOKEN" }
    }
  }
}

The hosted server also gives the agent this Actor's output schema, so it knows every field before the first call. apify mcp install <client> writes the config for you if you use the Apify CLI. Get a token at Apify Console → API & Integrations, and never commit it.

Tips for agents:

• "Summarise this PDF": pass the link in urls. Read text, and check status first: no_text and failed rows carry the reason in subStatus and error.
• "Find where the report says X": set includePages: true and quote pageNumber back to the user, or set chunkSize: 1500 and search the chunks.
• Long scanned documents: set pageRange or a lower maxOcrPagesPerPdf. OCR takes about 4–5 seconds a page at the default memory.
• The MCP call-actor tool waits up to 45 seconds. A longer run keeps going: fetch the results with get-dataset-items and the run's datasetId.
• Every row has the same keys and never a null, so an agent can read text, pageCount or title without a guard. An empty title means the PDF has no title in its metadata, not that the read failed.

Use it from code

from apify_client import ApifyClient

client = ApifyClient("<APIFY_TOKEN>")
run = client.actor("sumitr_mardy/pdf-text-extractor").call(run_input={
    "urls": ["https://www.irs.gov/pub/irs-pdf/f1040.pdf"],
})
for row in client.dataset(run["defaultDatasetId"]).iterate_items():
    print(row["status"], row["subStatus"], row["charCount"])

curl -X POST "https://api.apify.com/v2/acts/sumitr_mardy~pdf-text-extractor/run-sync-get-dataset-items?token=<APIFY_TOKEN>" \
  -H "Content-Type: application/json" \
  -d '{"urls": ["https://arxiv.org/abs/1706.03762"]}'