Refract for Windows Privacy Policy

Refract is a desktop text-rewriting app made by AZ Labs. It does not record your voice. When you invoke a rewrite, it reads text from the active field or selection using clipboard operations. It temporarily changes the clipboard and restores it afterwards.

Your text is sent to the AI provider you configure. If you configure backup providers, those providers may also receive the same text. Cloud providers process requests under their own terms and privacy policies. A local provider processes requests at the server address you configure.

New Windows setups use Refract Free by default. This route sends your rewrite input, instructions and any enabled context to the AZ Labs relay hosted on Cloudflare, which forwards the request through OpenRouter to a selected free-model provider. Availability and usage limits apply. You can choose your own provider or a local server in Settings instead.

Refract Free sends a random installation identifier and the client platform to the relay. Cloudflare supplies the connecting IP address. The relay stores keyed hashes of installation and IP identifiers, usage counters and request reservation timestamps to enforce usage limits. The relay implementation processes rewrite inputs and outputs in transient memory and does not log, cache or persist them. OpenRouter and its selected model provider receive the rewrite input under their own terms and privacy policies.

Depending on your privacy settings, Refract may include the active app name and window title to help choose a preset and guide the rewrite. Turn off this sharing in Settings if you do not want that context sent. Window titles and text can contain personal information; choose providers appropriate for your content.

When visible context capture is enabled, bounded text above the draft in the active window may also be sent to the rewrite provider. The optional screen-text mode uses local text recognition; screenshots remain in memory on your computer and only the recognized, bounded text is sent. Captured conversation context is excluded from rewrite history. Context and sharing controls are in Settings → Apps.

Optional Jev checks during rewriting are disabled by default. If you enable them and supply your own TypeSafe API key, Refract sends the original draft, candidate rewrite, selected preset instructions with applicable style rules, one-off instruction and contract version directly to TypeSafe. These checks exclude nearby conversations, app and window details, screenshots, and history examples. Text matching the app's sensitive-data patterns is skipped rather than sent. This filter cannot identify every sensitive detail, so enable checks only for content you are comfortable sharing with TypeSafe. Requests use your TypeSafe account and its usage charges and terms. ChatGPT sign-in does not provide TypeSafe access.

Jev returns experimental probability-based hints about meaning, tone and commitments before the rewrite appears in Review or is applied immediately. These hints do not establish factual accuracy. A check has a two-second deadline and no automatic retry. Failure lets the rewrite continue with a visible check-unavailable status.

Settings are saved in your Windows user profile. Saved API keys and authentication credentials are protected using Windows Data Protection for your user account. The TypeSafe key is stored in a dedicated encrypted file and is excluded from configuration exports. Authentication involves the provider's own sign-in service. Operational logs record provider/model metadata and errors. Error messages may contain information returned by a rewrite provider.

When edit learning is enabled, Refract keeps up to 150 rewrites and corrections in local encrypted history. It also retains up to 500 local Jev result and explicit feedback metadata entries, including contract/model version, probabilities, latency, token counts, fixed failure categories and helpful/wrong/overridden feedback. Jev metadata excludes draft text, preset text, one-off instructions and keys. Turning off edit learning stops recording new history and Jev metadata; previously retained files remain until removed.

Feedback you choose to send, with a diagnostics report only if you add it, goes to AZ Labs through the Refract relay and is tracked in Linear. The relay temporarily stores the report for delivery for up to 7 days, deletes it once delivery completes, and keeps content-free delivery and duplicate-prevention metadata for 30 days. A complete report also goes to the AZ Labs feedback email inbox, and a limited excerpt and issue link go to private Discord. Your optional contact email is included only in the email report, not copied into Linear or Discord. Refract sends feedback only when you press Send in the Send feedback window, which shows the exact diagnostics report before it goes. Diagnostics never contain API keys, sign-in tokens, your draft text, rewrites or error text returned by a provider. Refract adds a random submission UUID so retrying an unchanged report does not create another issue. The relay also receives the same random installation identifier and client platform as Refract Free and uses keyed hashes of them and of your IP address to limit repeated messages; they are not part of the report the AZ Labs team reads. If you choose Email instead, your own mail app sends the message.

You can change providers, remove credentials, disable Jev checks and change context-sharing settings in the app. Disabling Jev or removing its key cancels active checks. Local history's Clear history action confirms removal of rewrites, learned corrections and Jev metadata and cancels active Jev checks. Store uninstall removes the installed package; do not assume that it deletes every external settings or log file. Review and remove retained local settings if you need to erase them.

For privacy questions or support, contact AZ Labs at https://azlabs.ai/contact.