True redaction

Most "redaction" tools draw a black box over the text — the words are still in the file, one copy-paste away. POST /v1/pdf/redact physically removes the text objects under each region from the file, then draws the box. What was redacted cannot be recovered.

Request

Regions are normalized 0..1, origin top-left, per 1-based page:

PDF=$(base64 -w0 contract.pdf)     # macOS: base64 -i contract.pdf

curl -X POST https://api.pdf-forge.dev/v1/pdf/redact \
  -H "Authorization: Bearer $PDFFORGE_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{
    \"pdf\": \"$PDF\",
    \"regions\": [
      { \"page\": 1, \"x\": 0.1, \"y\": 0.22, \"width\": 0.35, \"height\": 0.03 }
    ]
  }" \
  --output redacted.pdf

const redacted = await client.pdf.redact(pdfBytes, [
  { page: 1, x: 0.1, y: 0.22, width: 0.35, height: 0.03 },
]);

The response is the redacted PDF; the X-Redacted-Objects header counts the text objects that were removed.

Find the coordinates with perceive

Don't compute coordinates by hand. POST /v1/pdf/perceive returns every text element with its bbox in the exact frame redact accepts, plus semantic labels (email, phone, iban, amount...):

const doc = await client.pdf.perceive(pdfBytes, { detail: "full" });

// Redact every element labelled as an email address
const regions = doc.pages.flatMap((p) =>
  p.elements
    .filter((el) => el.semantic === "email")
    .map((el) => ({
      page: p.page,
      x: el.bbox[0], y: el.bbox[1],
      width: el.bbox[2], height: el.bbox[3],
    })),
);

const redacted = await client.pdf.redact(pdfBytes, regions);

See Inspect for the full perception layer, and the MCP server whose redact_pdf tool wraps this whole loop (by id, regex pattern or semantic label) for AI agents.

Scanned PDFs: flatten after redacting

On a scanned page the text is pixels inside an image — the box is painted over it, but the pixels underneath survive in the image stream. Rasterize the redacted file (to-image at 200 dpi, then from-images) to make the removal irreversible. The MCP redact_pdf tool does this automatically with flatten: true.

Strip an overlay watermark

The reverse operation — removing a big diagonal "DRAFT" without leaving a white patch — is structural too. POST /v1/pdf/strip-watermark deletes the large rotated text blocks:

curl -X POST https://api.pdf-forge.dev/v1/pdf/strip-watermark \
  -H "Authorization: Bearer $PDFFORGE_API_KEY" \
  -H "Content-Type: application/json" \
  -d "{\"pdf\": \"$PDF\", \"minFont\": 36}" \
  --output clean.pdf

minFont (points) is the font-size threshold above which rotated text is considered a watermark.

const clean = await client.pdf.stripWatermark(pdfBytes, 36);

Next steps

Inspect — the perception layer that feeds redact its coordinates.
MCP server — agent-grade redaction by pattern or semantic label.
Secure — encrypt what remains.

Request

Find the coordinates with perceive

Strip an overlay watermark

Next steps

On this page