What shipped

ChatWrite confirmation (preview first, then Apply or yes in chat) now applies only in the ChatGPT and Claude apps. Automations and other connected clients write directly again, and every automated caller fails loudly if it ever gets a preview instead of a write .

i18nThe channel sign-in card can now be translated in every supported language. The sign-in card customers see when they connect from a messaging channel was stuck in English because its translation group was never registered. It is now, so the 14 translations can be published.

i18nThe language-mix check now treats OS and AI-client names as product names. Translation publishing works again after the two-step write change. The automatic translation publisher kept answering "published" while its writes were only previews, so new labels stayed in English until someone re-ran it. It now sends the confirmation and refuses a preview answer.

DocsThe docs overview map shows the Sign-in and identity page in your language. On the Arabic, Hindi, Japanese, Korean, Russian and Chinese docs overview, the page map listed the new Sign-in and identity page in English. It now has a short label in all 14 languages, and a test makes sure the next docs page cannot ship without one.

InboxG4: with a workspace's pii read masking ON, the Inbox thread translation and the email copilot (Console and MCP summarize_email_thread, suggest_email_reply, translate_email, and the Console copilot on an email conversation) now reads the mail thread masked, so drafts and summaries cannot quote raw card numbers, IBANs or national ids. With the setting off nothing changes.

InboxG4: with a workspace's pii read masking ON, the transcript Markdown export (Console and MCP) now masks the Handoff header row, and the MCP draft_tenant_support_reply copilot reads the conversation masked like the Console copilot does. With the setting off both are unchanged. MCP activity records name the workspace and the result size. Each call an AI assistant makes to Busymate AI is now recorded with the workspace it targeted and how big its answer was, in bytes and rows. When a ChatGPT or Claude view shows no data, the cause can be found from the activity record alone.

ConsoleThe script tag now says, once in the browser console, which call replaces an old one. Identify and refreshIdentity on the embed script keep working. The first call of each prints one console note naming the call to use instead (identityChanged), so integrations can move to the same names as the Web SDK (getIdentity, identityChanged, signedOut, account). The old v1 client drop-ins under /sdk/v1 are served unchanged and now answer with a Deprecation header.

ShopifyWhatsApp, Messenger and Telegram customers can sign in with your own login. A workspace can now let its chat customers sign in through the workspace's own login page. The customer sends "sign in", opens a one-time link, signs in, and sends the 6-digit code shown back in the same chat. Account tools then run as that customer with no extra card. The feature is off until the workspace lists the channels in channels.customerSignIn. ChatGPT and Claude: a write now asks first. The first call shows a preview (the card with Apply) and changes nothing; the change happens only when you press Apply or say yes in the chat . Signed-out shoppers get a reply again on stores whose site says "nobody is signed in". A signed-out shopper on a Shopify storefront (and any site whose getIdentity returns nothing) pressed Send and the chat stayed on "Still working…".

PlatformG4 per-workspace pii read masking, default off.

DocsThe Sign-in and identity docs page now has its navigation title in every supported language. The title of the new /docs/identity page showed in English in the docs menu of the Arabic, Hindi, Japanese, Korean, Russian and Chinese sites. It now ships with all 14 translations.

ChatBusymate AI clamps list sizes and pages instead of refusing. When an AI assistant asks Busymate AI for more rows than a list returns, for example 500 conversations at once, it now gets the maximum with a short note "limit clamped to 200" instead of an error. The conversation list also returns a next_cursor to fetch the next page. In ChatGPT, an error that arrives only as a code (such as "INVALID_ARGUMENT") now shows a sentence people can act on.

InboxDeleting a contact now also clears their name from the AI conversation records. When you delete a contact in the Inbox, the name and handle that the channel stamped on the contact's AI conversations is cleared too. The conversations stay; they just no longer show the redacted name.

PlatformThe login-page check on the sign-in provider test now actually runs. The "Test identified launch" report was meant to include a check of your login page (does it stay silent, can it be framed, does its cookie work in a frame), but it never appeared. It now does, with a fix text for each problem.

ChatA signed-in page no longer pops a guest chat open after a reload, and signing out ends the chat's account access. Some sites show the assistant inside one of their own pages (RentADriver's Support tab) and also have the floating chat button. On those sites the floating chat no longer reopens by itself, as a guest, every time the page is reloaded. After the visitor signs out, it can no longer pick up the signed-in session.

DocsOne page now explains every way to sign your customers in, and how your own tools act as them. The new page /docs/identity covers a customer already signed in on your site, sign-in inside the widget, the full chat page, apps, your MCP server or API, the "Ask customers" switch, sign-out, and what each channel can do. It starts with one diagram, then short numbered steps with copy-paste code, a verify checklist and a troubleshooting table. The older identity guides point to it.

ChatEdits to a sign-in provider (login URL, account URL, issuer) apply at the next chat load, with no republish. The chat page used the copy of the provider frozen in the last published settings, so a changed login URL only took effect after a publish. The published settings now only list which providers are on; the details are read from the live provider (cached for 10 s).

InboxIntegrators and agents can now read the Busymate AI identity contract as one machine-readable page. Https://busymate.ai/.well-known/Busymate AI-integration.json lists the identity contract version, the SDK releases, the launch and connector token rules, the host calls and the reason codes, each marked live, planned or legacy. Nothing changes in how sign-in or connectors behave. Messenger conversations taken over by a person keep the customer's Facebook name. When someone on your team claims a Messenger conversation, or the customer asks for a person, the Inbox now shows the customer's Facebook name instead of "Anonymous visitor".

MCPWorkspaces can stop asking signed-in customers to authorize account tools, connector by connector.

InboxPartner widgets in the chat can refresh their own read-only data again. Messenger conversations now show the customer's Facebook name. The Inbox, conversation header and Customer Dashboard name a Messenger customer by their Facebook name instead of "Anonymous visitor".

MCPConnectors can now receive the new signed customer token, one workspace at a time. When the new token is switched on for a workspace, every call a signed-in customer makes to one of its MCP servers also carries Busymate AI's ES256 identity token in the X-Busymate AI-Actor header (or in Authorization once the server says it accepts it), and guests' calls carry no identity token at all. The account-linking timeline shows when a connector accepted or rejected the token. Nothing changes for workspaces where it is off, which is every workspace by default.