API

Authentication

Send an API key as x-api-key. Keys are minted in the dashboard and carry their own per-minute rate limit, set by your plan. Anonymous calls work too, at 20 a day, so you can try the endpoint before signing up.

Browser clients can instead send the session bearer token from sign-in, which is what our own front end uses.

POST /v1/clean

curl https://api.removellmwatermark.com/v1/clean \
  -H "content-type: application/json" \
  -H "x-api-key: rlw_..." \
  -d '{
    "text": "Paste anything here.",
    "options": { "punctuation": false },
    "model": "claude/opus-5"
  }'

{
  "ok": true,
  "data": {
    "text": "Paste anything here.",
    "deleted": 3,
    "replaced": 1,
    "suspicion": 41,
    "verdict": "3 invisible characters found and removed. ...",
    "findings": [
      { "technique": "zero-width", "name": "ZERO WIDTH SPACE",
        "codePoint": "U+200B", "count": 3, "removed": true }
    ],
    "caveats": ["A statistical watermark ... cannot be removed by cleaning."],
    "usage": { "used": 12, "limit": -1, "plan": "developer" }
  }
}

Decoded payloads

Both endpoints return a payloads array. Hidden characters are not always just a marker: three schemes carry actual data, and each has a fixed published mapping, so a run of them can be read back rather than merely counted. Variation-selector byte encoding, Unicode tag sequences and zero-width binary are all decoded to text. A decoded payload puts the suspicion score at 100, because it is proof of deliberate insertion rather than a stylistic habit.

There is also a whitespace array for spacing used as a carrier: inconsistent double spaces after sentences, or trailing spaces on a suspicious share of lines. Those are reported as suspicions, since a careless human produces them too.

POST /v1/detect

The same scan with no rewritten output, the report only. Use it when you want to know whether a document is marked without changing it.

Limits and errors

GET /v1/limits reports what the caller may do right now. Over the per-minute ceiling returns 429; over the daily quota or off-plan returns 402 with an upgradeUrl. Every response carries ok, so a client never has to guess whether it holds data or an error.

What we store

Counts and code-point classes, for our own usage numbers. Not the text, not a fragment of it, not a hash of it. Scan rows are deleted after 90 days.