What happens to what you paste.

The text you paste or upload is sent to the model once, with the page text, to produce the analysis. It is held in memory for the length of the scan and is not written to storage. The report records the document's file name and its character count, and each finding quotes the sentence from your document that the page drifted from. Nothing else from the document is kept.

While you are filling in the form, an unsent draft is kept in your own browser's local storage so a refresh does not lose it. It is removed when the scan completes. It never leaves your browser.

Each page you list is fetched once with a plain HTTP request from our server, identified as msgdrift/1.0. Scripts are not run, nothing is logged in, and only public hosts are fetched. The page text is used for the scan and discarded. Findings quote the sentence from the page that drifted.

The report is stored so its link keeps working. It contains the site name, the page URLs, the scores, the summary, and the findings with their quoted sentences. The link is a random identifier that is not listed anywhere and is marked not to be indexed. Anyone who has the link can read the report, so share it the way you would share a document.

The browser that ran the scan receives a delete token. From that browser, or from the delete link it can copy, the report can be removed at any time. Deletion is immediate and permanent. Only a hash of the token is stored with the report, so we cannot recover a lost delete link, and we cannot delete a report on your behalf without it.

Analysis runs on Anthropic's Claude models through Vercel AI Gateway. The gateway does not retain prompt or response content after a request completes; it keeps request metadata such as model, token counts, and duration for up to 30 days. Every request from msgdrift asks the gateway for zero data retention on the provider path, which routes only to endpoints that do not store prompts or responses. Vercel does not train on these requests.

Scans are limited per visitor so the shared model budget lasts. To count them, your IP address is combined with a secret and hashed, and the hash is kept for at most 24 hours. The address itself is not stored. Beyond the hosting platform's standard request logs, which we do not inspect, no other information about you is collected. There are no analytics scripts, no cookies, and no accounts.

If you would rather nothing leave your machine except the model call, the command line version runs the same analysis with your own Anthropic key and writes the report to your disk. The web app's source is public at github.com/wpsands/msgdrift, so every claim on this page can be checked against the code.

msgdrift is made by Page Sands. Write to page@sandsdx.com. If this page changes, the date at the top changes with it.

Scan a site