Extension data and permissions
beforeword helps you examine selected or pasted writing. Hovering and selecting alone send nothing to a model. The toolbar, context menu and beforeword symbol capture or display text without a model request. To send the displayed passage, choose “Examine writing” in the panel or press Ctrl/Cmd+Enter in its text field. A separate dictionary action sends the chosen word to Wiktionary.
Chrome extension 0.5.6 · Updated 15 October 2026. This policy covers the extension, including its optional AI and dictionary features. The main website has a separate data policy.
Purpose and your choice
Text is processed to show the writing, offer a reading beside exact excerpts, or look up a chosen word. The destination and the passage to send are shown in the panel before analysis. You choose whether to start analysis with the button or keyboard shortcut, which connection to use and whether to enable the in-page button for a site. Prepared examples do not call a model.
In your browser
The latest captured passage, its site origin without the path or query, and tab and passage-position information use temporary extension storage until replaced, cleared or the browser session ends. Where session storage is unavailable, these data remain in memory and can disappear when their extension context closes.
A temporary beforeword token is created for an explicit introductory analysis and expires on the server after 24 hours. A separate random installation identifier persists locally between sessions to count the 8 introductory analyses. It is not an account name.
Language, enabled site origins, model connection settings without the key and the cached allowance persist locally. A confirmed own-model checklist also saves its policy version, connection profile, accepted items, confirmation time and a SHA-256 digest binding the settings to the key; it does not save the key itself in permanent preferences. Current text, reports and dictionary results stay in panel memory. Hiding or closing the panel does not guarantee that its memory is destroyed. beforeword does not synchronize these settings through Chrome Sync.
A personal API key, including one obtained through OpenRouter sign-in, uses temporary extension storage or memory. It is not saved in permanent preferences, exposed to page scripts or sent to beforeword’s server. A browser session can contain several openings of the panel.
When text is sent
The introductory AI connection sends data to api.beforeword.xyz. Session creation sends the installation identifier and quota-version information, without the selected text. Analysis sends the selected passage, language and format information with a temporary session token. The server adds the beforeword instruction and sends the model request to OpenAI.
The source page’s address is not added to the model request. A URL or personal information written inside the passage remains part of the passage. Installation identifiers and session tokens are not forwarded to OpenAI as separate fields. Network requests also expose ordinary connection metadata, including an IP address, to their recipients.
OpenAI in the introductory connection
OpenAI says API data is not used for training unless the customer opts in. Its abuse-monitoring logs may contain prompts and responses, kept by default for up to 30 days, with legal and safety exceptions. beforeword sends store:false; this does not promise zero retention.
Your own model
In the optional own-model connection, the passage, language, beforeword instruction, model choice and your API key go directly to the service you select. An Anthropic workspace identifier is included if configured. These requests do not pass through beforeword’s server. Loading the available model list sends the key but no selected passage.
The connection list includes OpenAI, Anthropic, Google Gemini, xAI, OpenRouter, DeepSeek, Mistral, Groq and a compatible endpoint you enter. Availability of text analysis depends on the processing mode described below. OpenRouter can forward a request to another model provider. The selected service, route, account and plan determine retention, permitted security review and charges.
Own-model processing modes
Before sending a passage to your own model, the extension requires you to review the connection-specific conditions and explicitly confirm the applicable account settings. It does not read or independently verify your provider account settings. A change of provider, protocol, address, model, workspace or key requires a new confirmation. Obtaining a model list sends the key to the chosen service and can happen before this review; it sends no passage. The extension does not automatically switch to another provider or relax these request conditions after a failure.
For OpenAI, Anthropic, xAI and Groq, the review requires that the account has not enabled or authorized general model training on this material. These API services describe no general training by default or without customer permission. That does not mean no retention or human access: their policies allow operational, abuse-prevention and legal processing. Standard periods and exceptions differ by provider, model and agreement; approved zero-retention arrangements are separate. beforeword does not promise that such an arrangement is active.
For OpenRouter, every analysis request requires upstream endpoints classified by OpenRouter as denying data collection and supporting zero data retention. The account review separately requires prompt/response logging, data-sharing discounts or model-improvement sharing, and Broadcast/export of this material to be disabled. Request routing cannot turn those account options off. OpenRouter can still keep request metadata and permits some temporary memory caching under its zero-retention classification; this is not a promise that every record throughout the chain is immediately erased.
Gemini analysis requires your confirmation that the key belongs to a Cloud project with active billing, that this material is not contributed through shared datasets or improvement feedback, and that you meet its age, region and professional-use requirements. Unpaid processing is not offered for this feature. Paid processing still includes abuse-monitoring retention, ordinarily 55 days, and possible review of flagged content under Google’s terms.
Mistral analysis requires Anonymous improvement data to be disabled in Admin → Privacy. This release allows only the reviewed generally available model IDs mistral-small-2603, mistral-medium-3-5 and mistral-large-2512. Moving aliases and Experimental, Labs or Preview models are not enabled: their terms can permit training despite other settings. Ordinary API content can be retained for up to 30 rolling days for abuse monitoring; approved account-level zero retention is separate.
DeepSeek remains in the connection list and can load its model list. Direct text analysis is disabled in this release because a documented account-wide API mode excluding general training on all supplied material has not been established. A personal-data opt-out does not by itself establish that mode. No passage is sent to the direct DeepSeek API for analysis.
For a remote compatible endpoint, you must have checked that its operator and downstream processors use the material only for the requested feature and related security, and exclude general training or unrelated reuse. Recognized official provider hosts retain their provider-specific checks. For localhost or 127.0.0.1, you must have checked that the server processes locally without external forwarding or general training. A loopback address alone does not prove this. These are your declarations about the actual service, not a technical verification or privacy certification by beforeword.
Connect an account
OpenRouter sign-in is optional. After your action and approval on its authorization page, the extension exchanges a temporary authorization code and verification data for an API key. It does not read chat cookies or use your Google profile. The key follows the temporary storage rules above. Other model connections use an API key you provide.
Online dictionary requests
The selected query goes directly to en.wiktionary.org/w/api.php or ru.wiktionary.org/w/api.php. The whole passage, page address, beforeword ID and model key are not included. Requests omit credentials and referrer; Wikimedia still receives the query and ordinary network information, including the connection address.
Definitions, lookup results and the current path use page memory, with no permanent dictionary history added by beforeword. Destroying or reloading the panel clears that memory; hiding the panel may keep it alive. The browser and Wikimedia can handle network data under their own policies.
Definition text is extracted from a linked Wiktionary revision and credited to its contributors under CC BY-SA 4.0. Examples, references and formatting are removed; HTML spacing is simplified without paraphrasing. The source entry remains available for its full layout and notices.
beforeword’s AI server
The beforeword AI application does not save or log the passage or report. Its quota database stores hashes of installation identifiers and session tokens, a network identifier derived from the connection address with a server secret, the extension origin, expiry times, counters and temporary reservations. These records operate the allowance, prevent concurrent overuse and protect the service. A hash is not a guarantee of anonymity.
An installation’s completed-analysis count has no daily reset or automatic expiry. Sessions expire after 24 hours; temporary reservations expire after 2 minutes. Expired temporary records are removed during later database operations. Resetting the panel does not remove these server records or backups.
Hosting and service records
beforeword’s API runs on Namecheap VPS hosting. Technical web-server logs may contain the connection IP address, date and time, request path, response status and browser information. Non-empty logs are configured for daily rotation with up to 14 archives. These settings do not guarantee deletion from all backups or the hosting provider’s own records; those are separate from the application’s quota database.
The extension adds no analytics, browser fingerprinting, advertising, external fonts or remotely loaded executable code. This statement describes beforeword’s extension, not every service or external website you choose to use.
Permissions
Site access and personal-model host access are optional and requested for the chosen origins. The broad optional host patterns in the package allow you to choose a site or endpoint; they do not grant every site automatically. Chrome requests identity permission when you connect OpenRouter. Turning off the in-page button does not revoke a browser permission already granted; manage it in Chrome’s extension settings.
External model and dictionary requests use HTTPS. A manually configured model on localhost or 127.0.0.1 may use HTTP on your own computer. The extension does not automatically read all tabs or your browsing history, and excludes editable fields from page capture.
Reset
Reset clears the current passage and report, captured selection and local beforeword token. It preserves the installation identifier, allowance and own-model settings and key. Use the separate forget-key action to remove the locally held personal key and its saved review confirmation. A key created through OpenRouter sign-in can also be revoked in the provider account. Browser extension settings control granted site access. Copied files and data already sent to recipients are separate.
Limited Use commitment
beforeword’s use of extension user data adheres to the Chrome Web Store User Data Policy, including its Limited Use requirements. beforeword uses and transfers data only for the disclosed reading features and their necessary operation and security. beforeword does not sell user data or use it for personalized advertising, data brokerage or credit decisions.
beforeword permits human access only with your explicit consent to specific data, where necessary for security or law, or for lawful internal operations using aggregated and anonymized data. This commitment governs beforeword’s handling; the linked recipients describe their own processing.
Chrome Web Store · Limited Use
Recipient policies
Only the recipient needed for the action you choose receives that request. For a custom compatible endpoint, consult the operator of the address you enter. Provider terms can change; the links below describe the services separately.
- OpenAI
- Anthropic
- Google Gemini API
- Google Gemini API · Logging
- xAI · API
- OpenRouter · Routing
- OpenRouter · Data
- OpenRouter · ZDR
- DeepSeek
- Mistral · Privacy
- Mistral · Terms
- Mistral · Opt-out
- Groq · API data
- Groq · API terms
- Wikimedia
- Namecheap
Contact and deletion requests
For questions about this policy or a request concerning data held by beforeword, contact mail@beforeword.xyz. Describe the feature and approximate time involved; do not email API keys or session tokens. For data held by your chosen model provider or Wikimedia, use that recipient’s privacy contact. Changes to this policy will be dated on this page.