claudewatermarker.

Privacy, by processing mode

Light

Light mode never leaves your browser.

Character operations, positions, visualization and diff are computed locally. Drafts and reports live in page memory and are discarded when you refresh or close it. Light does not create a device identifier or send usage events. Deep creates a random device UUID in localStorage for quota checks; no draft is saved there. A future paid license may also be stored locally.

Deep

Deep rewrites your wording and is processed on our servers. The server algorithm uses a third-party processing service. This application does not save request text, response text, or text-bearing logs. We cannot promise that processing or hosting providers have no retention; review their policies before sending sensitive text.

The working draft and recent duplicate results stay in page memory so you can retry or reuse a run. Server protection records contain HMAC identifiers, job status, counters, tokens and financial exposure. They expire automatically; they do not contain submitted text or results. The server code does not write input or output to a database, file, analytics event, cache, or application log. Network providers and DeepSeek operate under their own policies; this is not a promise about their retention or training practices. These modes remain gated until deployment and provider review are complete.

Anonymous free allowances

Each browser identifier has 5 free qualifying Deep runs per UTC day. Each public IP address shares a total of 5 anonymous runs per UTC day, including reserved requests in progress. Private windows may create a new identifier but still share this network allowance. People on the same household, workplace, school or VPN connection may share the cap. This does not identify a physical computer. IP addresses are used to derive a server-side HMAC quota key; this application does not store the raw IP in quota records or collect browser fingerprints. Limits reset at 00:00 UTC. Quality previews and eligible failed attempts restore their reserved quota; actual processing costs remain recorded.

Meaning comparison

When enabled, Deep sends the original wording and valid rewritten candidates to Cloudflare Workers AI for BGE-M3 embedding comparison. Vectors are compared in the Worker and discarded after processing. Similarity is not a percentage of meaning preserved, a fact check, or a watermark measurement. Cloudflare operates under its own data-processing policies; this application does not save draft text or vectors.

Controlled bench experiments

Submitting a prompt on the verification page is a separate, explicit server action. We generate a new sample with our own SynthID-Text keys. The original words are not retroactively watermarked. Bench generation and embedding inference run on the configured bench server; an edited sample is sent to DeepSeek in Standard or Deep. We do not offer an endpoint that classifies arbitrary text as AI-generated.

Clipboard and analytics

Copy writes the result to your device's clipboard; operating systems and extensions may retain it. Optional analytics records only numeric metrics, booleans, and allowlisted mode/error enums. Text, tokens, characters, URLs inside drafts and provider payloads are excluded.

Infrastructure and contact

Ordinary page requests and provider operations may generate infrastructure metadata. Synthetic development fixtures are explicitly labeled, contain no submitted visitor text, and may be saved as reproducibility artifacts.

DeepSeek privacy policy ↗ · Return to the tool