Chrome extension audits, prompts and chat

The Chrome extension foundation lets you inspect a registered website from a side panel. It is available as a local development build; a Chrome Web Store release is still pending. Optional toolbar tools prepare prompts for supported AI chat apps; live-host approval is still pending. The separate Tailwin Chat view uses workspace model allowances and explicit per-turn review.

Link your workspace

A workspace owner or admin opens Settings and creates a labelled extension token. Copy the secret when it appears; it will not be shown again. Open the extension and paste it into the link form. The panel loads your workspace brand and active sites. Tokens expire after 90 days, and a workspace can have five active tokens.

Revoke a token in Settings to stop future requests from that link. Unlinking clears the token from this browser. Removed or demoted managers lose token access. The token is stored in trusted local extension storage, not Chrome sync, and is never sent to the page you inspect.

Choose an audit

Open a page on your registered HTTPS site, invoke the extension and choose the matching site. Opening the panel does not start an audit. Pages with query parameters, credentials, unsupported ports or a different site hostname are excluded. AI chat hosts are excluded from all current-tab audits, so a conversation tab cannot become the audit target.

URL audit sends only the chosen URL. Tailwin fetches the public page and returns citability, readability and structured-data findings. A logged-in browser page does not make that content publicly accessible to Tailwin.

Rendered-page audit is a separate action for an authenticated or unpublished page on the registered HTTPS host. The extension prepares a sanitized capture, shows a review summary and asks you to confirm upload. Forms, editable/hidden content and executable code are removed, but visible text can still contain private information. Review it before sending. If the page changes, capture it again.

Rendered findings are imported browser evidence. They cannot prove what a crawler receives, its HTTP headers or the speed of your logged-in page. Raw submitted HTML is not retained. Rendered evidence stays separate from measured server-fetch history.

PageSpeed uses the public URL and requires separate quota confirmation. Choose mobile or desktop. Lighthouse lab results and CrUX field data are different evidence; missing field data is unavailable, not zero. Cached results and typed failure receipts are labelled. Review a pending or failed receipt before trying again.

Access and limitations

Chrome grants page access when you invoke the extension. Navigating to another origin or closing the tab can remove that access; invoke it again on the intended page. The extension does not automatically capture or transmit when you switch tabs, and does not automatically retry quota actions.

Partner logos may load from the configured HTTPS image host without a referrer or extension token. A failed logo uses a fallback. Default audit-only linking does not enable paid keyword research or chat. Toolbar research requires an explicitly selected capability and a separate action confirmation; the toolbar never creates a Tailwin Content draft.

Use the Signal guide for evidence interpretation and integrations and billing for workspace access and metered services.

Prepare prompts for your chat app

Toolbar access is separate from audit access. In Settings, create a new extension token and explicitly select stored evidence and shared settings capabilities. Select competitor research only if you intend to request it. Existing audit-only tokens stay audit-only; a missing toolbar capability does not unlink a valid audit token.

In the extension panel, enable each supported chat host separately. Disabling a host removes its permission and toolbar. An adapter marked unverified or disabled remains unavailable; you can still prepare and copy a prompt from the panel. The proposed host list is ChatGPT, Claude, Gemini, Perplexity, Microsoft Copilot, Grok, DeepSeek and Mistral Le Chat. All production adapters currently remain disabled pending terms and compatibility review. Local fixtures do not establish permission to integrate with, or live compatibility with, these services.

The composer toolbar offers AI Search Review, Schema, PageSpeed, Outrank, Fix Brief, Content Strategy and Blog Draft. Its settings popover carries model preference, language, locale, market, tone, output format, depth, live-data preference and target site. Model preference supplies prompt context; it does not switch the chat site's selected model. Preferences are shared with your Tailwin user settings.

Choose a tool, then review the owned target page and settings in the extension panel. Approve the evidence request before Tailwin retrieves private workspace evidence. Stored-only mode starts no page fetch, scan or provider request. Fresh evidence also requires live data on and an explicit fresh selection. Fresh PageSpeed consumes Google quota; fresh competitor research requires its separate disclosed cost confirmation. No operation is retried automatically.

Review the prepared prompt and evidence dates, sources and availability. Choose Insert without sending or Copy prompt. Insertion preserves your existing unsent draft without reading or backing it up. You decide when to press Send. If the composer, document, permissions or settings change, prepare a new bound action or use the selectable copy fallback. The eventual generation uses your chat subscription and does not incur Tailwin model spend on this toolbar path.

The toolbar does not read, retain or transmit host conversation content or drafts. Evidence you explicitly insert becomes visible to the chat site. Adapter diagnostics contain only the supported host, adapter version and failure category. Production privacy and infrastructure logging review remains required before distribution.

Chat inside Tailwin

Create a token with the optional Tailwin side-panel chat capability. Include shared settings capabilities to read/save preferences and evidence access for tools. Old tokens gain no access automatically. Open Chat and load the current settings. Unavailable models remain visible; select an allowed, configured model before sending.

Type or paste your own message, choose Review generation cost, then confirm the reviewed turn. The quote shows model, provider, conservative estimate and output limit. The final receipt separates quoted, reserved and settled amounts; the estimate is not a guaranteed invoice ceiling. Model-owned search may carry the catalog's request fee independently of the live Tailwin data toggle. Mock AI is visibly labelled and has zero model cost.

The panel does not import host conversations or drafts. Your Tailwin messages and answers remain in panel memory and disappear when it closes or you choose New chat. Only operation identifiers are retained temporarily in trusted session storage for recovery. The server stores bounded operation/receipt metadata and expiring tool projections, never the chat transcript. A lost answer cannot be recovered or regenerated automatically.

Use Reviewed tools to open any of the same eight MCP actions without paying for a model proposal. Enter the selected site's arguments, review the exact action and approve execution. A model suggestion is untrusted and executes nothing. JSON-LD operator facts must be supplied and reviewed by you; model-generated claims are never attestations. Fresh tools require live data and the relevant capability; quota/paid work requires its separate confirmation. Tool completion never starts another model turn. Choose Use this evidence in a new model turn and review a new quote if you want an explanation.

If a request times out, use Recover status. Pending or uncertain work is not retried, and a generic unavailable response does not prove nothing was charged. An administrator may need to reconcile the retained receipt before new work can proceed. The application permits one active operation per token and four per workspace, including mock/free actions. These are application limits in addition to provider and workspace spending limits.