When you ask an AI agent to redact a scanned page and text redaction finds nothing to replace, the page is not clean, it has no text layer to search. To redact a scan locally, the agent covers a region by coordinates with redact_image_area instead of matching words. With GroupDocs.Redaction.Mcp connected to Claude Desktop or Cursor, the request looks like this:

Cover the bottom third of scan.png with a black box.

The step-by-step version with config and troubleshooting is in the documentation: How to redact scanned documents and images with an AI agent.

Why does text redaction find nothing on a scan?

A scan is a picture of a document. The words on it are pixels, so a regular expression has nothing to match. redact_text finds nothing to replace on a scanned page, and finding nothing is not the same as nothing sensitive. Treat an empty outcome on an image-only file as “no text layer”, then switch tools. The same applies to photographs, stamps and a scanned signature page at the back of an otherwise digital contract.

Which tool covers a region of a page?

redact_image_area covers a rectangle of a page with a solid-colour box and permanently hides the image content underneath, such as faces, signatures and stamps. It takes x, y, width and height in pixels, measured from the top-left corner of the page, and an optional color (default Black, or a hex value such as #FF0000). The tool takes no page argument, only the file and the rectangle, so open the output and confirm where the box landed. Because the numbers are pixels, you need the image or page size before you choose a rectangle.

How do you get the coordinates?

Start with the page size. get_document_info returns the file type, page count and per-page width and height, which turns “the bottom third” into real numbers.

How big is scan.png?

Then give the agent the region:

Cover the bottom third of scan.png with a black box.

For a specific element such as a signature or a face, you need its position. Two practical routes work. Open the page in any viewer and read the coordinates yourself. Or let the agent render the page with a server that returns page images: the GroupDocs.Viewer.Mcp render_page tool renders a single page as a PNG and returns it inline, and the GroupDocs.Annotation.Mcp server can return page previews (its preview tool currently fails for PDF files on Linux and in Docker; Word and other Office documents render, and on Windows with dnx every format renders). A vision-capable model can then estimate the box, and you confirm it before the redaction runs.

What about documents that are half text, half scan?

A born-digital contract with a scanned signature page needs both tools, and redact_image_area cannot be aimed at one page by an argument. Keep the pieces separate: redact the text first, and treat the scanned pages as images that get their own area redaction.

Redact the email addresses in contract.docx, then tell me which saved file I should use for the next step.

Then check every page of the final output by eye.

What does a typical scan workflow look like?

The order that works is: read the page size, decide the region, cover it, then look at the result. If the region is not obvious, render the page first and let a vision-capable model propose a box. Keep the proposal and the final coordinates in the transcript so that anyone reviewing the case file can see what was covered and why. Redacting by region is blunt on purpose: it covers a rectangle and nothing else.

Does this server read text in images (OCR)?

No. GroupDocs.Redaction.Mcp redacts; it does not recognise text in images. That keeps behaviour predictable, because no confidence threshold is quietly deciding that a blurry name was not a name. If your workflow needs the text out of scans, extract it first with a tool built for that, decide what is sensitive, then redact by area. The server cannot tell you where a name sits in a picture.

Example session (abridged)

You:    How big is scan.png?
Agent:  -> get_document_info(file: scan.png)
        The image is 1200 x 1600 pixels.
You:    Cover the bottom third of scan.png with a black box.
Agent:  -> redact_image_area(file: scan.png, x: 0, y: 1067, width: 1200, height: 533, color: Black)
        Saved the output file. Please open it and check that the box covers the area.

How do you check an area redaction?

Look at every page of the output, not only the one you aimed at. An area redaction that is 20 pixels short is a redaction that failed, and no response text will tell you. This is the one case where verification cannot be automated by re-running a pattern: the check is your eyes, or a rendered preview the agent can see. Verification is covered in its own post, Trust no redaction you haven’t verified.

What are the limits of evaluation mode?

Without a license the server allows one document per process and one redaction, capped at 4 replacements, plus trial badges. For scans that means one box: a page that needs three covered regions gets one, and still looks processed. Run get_license_status first. The cap is for trying the tools, not for real work.

  • PDF on Linux. On Linux, including the Docker image, redact_image_area and erase_metadata currently fail on PDF files (the PDF engine’s image handling depends on System.Drawing, which .NET supports only on Windows); both work on Word documents, and redact_text works on PDF. Run PDF area redaction and metadata erasure on Windows with dnx.

FAQ

Can an AI agent redact a scanned document? Yes, by area. It covers a rectangle you define with redact_image_area, because there is no text for pattern matching to find.

Can Claude remove a signature from a scanned image? Claude can drive the redaction once the position of the signature is known, from a viewer or a rendered page preview. The server then covers that rectangle with a solid box.

How do I black out a region in a scan or image automatically? Give the agent the file and the rectangle, and it calls redact_image_area. The tool has no page argument, so check every page of the result by eye.

Why does text redaction change nothing on my scan? The scan or image holds pixels, not text, so there is nothing to replace. Use area redaction, and check the result by eye.

Go deeper