What shipped

DocsPublic preview knowledge reads. After a site's readonly documentation search returns a structured link, the assistant can read that exact document URL or its same-origin path and query in the same conversation. Successful page and preview-browser reads now carry short-lived server-signed evidence bound to the runtime session and source. Prose, failed results, foreign origins, altered receipts and other sessions do not grant evidence; current policy and approval gates still apply before every execution. Validation: actual page resolver → executor → frame result → same-turn re-resolution → read dispatch, plus policy-change and receipt forgery/replay negatives. The actual public preview roster is now observable without arguments or response data. Fresh same-origin remote knowledge capabilities take precedence over matching preview-browser tools; a failed knowledge candidate is temporarily removed only for the same session/source and its browser fallback returns. Host-page account/navigation and unrelated tools remain available. Natural live selection and widget acceptance are verified after the protected release.

DocsPreview MCP tool names. Public website previews retain the provider's original tool name and readable title, so a requested documentation tool is recognizable to the assistant and its result has a clear label. Internal identifiers, source integrity and read-only access checks stay intact. Tools without a returned label use a localized public-tool label. Validation: captured anonymous documentation catalog; 23 site-MCP runtime tests and 80 browser/render/disclosure tests; hosted qualification follows. The legacy page-tool path provenance gap is separate.

Docs2898 — Keep historical reference originals outside the runtime image. The image tar stream excludes only the release's reference-assets/archive subtree. An explicit nul-delimited file list and nonrecursive tar preserve exact path boundaries, hidden files and symlinks on gnu and bsd tar. Notes, docs, tests, maps and dependencies remain; some are read by runtime/gate tools. Packing leaves the original host release untouched. Controlled comparison of the actual build 1541 archive verified all 91,425 retained paths, file bytes, modes and link targets, plus 13 runtime/config paths; only 76,920,750 bytes of historical archive originals were excluded. Normal protected prebuild/app gates and the four deployment rollouts provide final runtime acceptance.

ConsoleKnowledge Index pause now stops fetching while keeping existing indexed content available. Console and MCP share the same pause/resume action. Scheduled and manual requeue respect the pause, with ordinary sync timing restored on resume. Keep Knowledge scheduling controls within the existing panel budget by loading provider credentials, catalogue choices and re-read options only after that provider is selected. The default website form, validation and controlled field values are preserved., . Workspace picker labels. Long workspace offers now truncate inside their row instead of hiding the workspace name or overflowing the picker. The full workspace name, offer, plan and slug remain available as text and tooltips; the selected workspace chip uses the same bounded badge.

ConsoleWorkspace connection pages now show only the selected workspace's integrations. Personal messaging sign-ins and their disconnect controls remain in Account Settings. Connected apps now lives in personal Account Settings; the former Console link opens that same Settings portal and preserves account-wide OAuth grant controls. 2895 — Bound dedicated builder cache after image publication. The dedicated rootless image build now shares the canonical workspace/maintenance lock and retains 20 GB of unused build cache after a confirmed push. Cache cleanup failure reports an already-published digest and exits 78, allowing the captain to resume with the existing immutable image. The fallback local image build no longer silently prunes a shared builder. Permanent 20 GB daemon GC was protected-published and live-verified separately in infra PRs 334/335; release archive slimming remains a separate audit.

MCPMCP Apps now receive complete arguments and standard tool-result envelopes, including the inner result of a successful connector call. Quote cards using the MCP Apps SDK can read the completed quote, while existing MCP results, errors and first-party widgets keep their payloads. . Public website MCP widgets can render in a /try preview through the existing sandboxed App host. Resource reads and callbacks are bound to the runtime-selected scan and authenticated visitor session; callbacks remain anonymous, read-only, explicitly app-visible and schema-checked. Source receipts expire after ten minutes and never authorize a call without visitor identity. Public website previews discover explicitly anonymous read-only MCP tools, keep business tools distinct from knowledge coverage, and render supported MCP Apps through the shared host. Catalogs, resources and callbacks remain bounded and bound to the authenticated preview session; private and write tools stay closed. MCP Apps now receive complete arguments and standard tool-result envelopes, including the inner result of a successful connector call. Quote cards using the MCP Apps SDK can read the completed quote, while existing MCP results, errors and first-party widgets keep their payloads. .

PlatformBusymate AI website analytics now uses its own Google Analytics property, separate from DevTools. Existing consent choices and exclusions for customer sites and native apps remain unchanged. Diagnostics release the Run control when a stalled check expires, including after an early timer wakeup. The card continues to receive server changes through realtime events.

DocsThe assistant can open an exact docs URL returned by a verified public knowledge search without asking you to paste it again. Successful search evidence is bound to the same connector and cannot authorize writes or supply customer identity details. MCP App initialization compatibility. Tool cards now receive sandbox permissions as an object and the tool name/input schema from the connector catalog. Both chat and the Console explorer retain the catalog definition. Older resources without a valid catalog schema omit the optional toolInfo field instead of sending an invalid tool descriptor. This fixes SDK initialization rejection seen on a successful RentADriver quote while preserving capability grants and sandbox restrictions. The assistant can open an exact docs URL returned by a verified public knowledge search without asking you to paste it again. Successful search evidence is bound to the same connector and cannot authorize writes or supply customer identity details. Delivery and other location-based tools now preserve the street and postcode details you supplied across turns. The assistant checks nested address fields before dispatch and corrects or clarifies ungrounded locations rather than inventing a city or changing postcode characters.

ConsoleSaving a verified MCP connector in the Console now applies its runtime assignment automatically. Editing an already selected connector also refreshes its published configuration. Refused checks leave the previous runtime live and show the normal save error. Private MCP connectors can pass publication readiness with verified customer identity without declaring any tool public. Public tools still require the public-read check, and failed identity, OAuth, or connector readiness keeps publication blocked.

MCPKnowledge base shows when a source uses MCP knowledge tools and automatic crawling is paused. Previously indexed content is retained. Reopen saved AI-only web conversations even after the visitor session expires. Staff authorization is checked separately; visitor expiry and revocation remain enforced for visitors, and native-channel admission is unchanged. Avoid a redundant reopen after staff promotion already opened the conversation. Fixes .

MCPBusymate AI MCP widget diagnostics and connection refresh are available on the live Edge service.

DocsWebsite-owner guide: knowledge through MCP. Website owners can implement read-only knowledge tools that search their own index and return citation URLs. The new guide includes TypeScript, Python and Cloudflare Worker examples, discovery files, public-tool security, and the existing direct MCP connection and guarded publication workflow. A new MCP-only setup does not need a Website crawl source.

VoiceBusymate AI widgets in ChatGPT show their data as soon as ChatGPT provides it. Widgets no longer wait for a slow handshake before showing a result ChatGPT has already handed over, and a loading bar replaces the empty white frame while a widget starts. Busymate AI widgets in ChatGPT carry the iPhone delivery fix and diagnostics. The Busymate AI MCP server now serves the widget bundles that keep early tool results and stop hidden refresh errors (busymate-ai-src). Support audit recovery. Support audits retry an incomplete or invalid analysis response once with a bounded larger output allowance. If analysis still fails after conversations were read, the audit keeps the scanned conversation and message counts instead of showing zero. New tools on a connected MCP server now reach your visitors automatically, and only if they really work without signing in. When an MCP server connected to your workspace gains a new read-only tool that works without signing in (for example a new docs search), the nightly connector refresh now offers it to your visitors on its own; before, it stayed hidden until someone classified it by hand. Servers where each customer signs in themselves are refreshed too. A tool that turns out to need sign-in is no longer offered to anonymous visitors, so your mate stops trying tools that can only fail. Identity provider assignments now save automatically in the Console. Saving an existing provider also applies its assignment when an older draft had never been published. A failed identity check keeps the current published configuration and shows the existing save error. White-label help sites no longer advertise the platform's management MCP server as their own. A white-label site (for example a workspace's help host) used to publish /.well-known/mcp.json naming the platform's own management MCP server, which signs callers in to their Busymate AI workspace and is not that site's MCP. An AI agent discovering the site's MCP read it wrong. White-label hosts now answer that address with a clean "not served", matching their agents.json, and their API catalog no longer lists it. busymate.ai itself is unchanged. Your mate looks things up with a site's own docs tools, and cites them. When a workspace is connected to a site's MCP server, your mate now treats that server's docs, help, search and setup tools as the site's knowledge. It answers questions about the product with them before saying it doesn't know, and cites the page links they return, just as it cites pages it has read itself.

PlatformVisitor events in Telegram now show the search keywords for their landing page.

ConsoleYou can now see and disconnect the apps connected to your account. A new Console page, Connected apps, lists every app you approved to use your Busymate AI account, such as Claude, ChatGPT or an app on your computer. For each one it shows whether it is a known app, the address its access goes to, which workspace it can use, and when it was approved and last used. Disconnect cuts its access at once. You can do the same from any MCP client with the new list and disconnect tools.

ConsoleConnecting Claude, ChatGPT and other MCP apps is safer. When an app asks to connect to your Busymate AI account, the approval screen now marks Claude, ChatGPT and VS Code as known apps and names who runs them. For any other app it shows an "Unverified app" warning next to the exact address that would receive access. An app can no longer register under a name like "Claude", "ChatGPT" or "Busymate AI" unless it really belongs to them. Unused app registrations are cleaned up automatically. A failed preview insight now says what went wrong and has a Retry button. In Console › Platform › Insights › Previews, a session whose AI insight could not be generated now explains in plain words whether the conversation could not be read or the AI models could not produce a valid insight. A Retry button re-queues it, and the card then shows "Retry queued" until the new insight arrives (about 5 minutes). The raw technical reason is still there under "Technical detail". The same action is available to MCP clients as retry_preview_insight. Preview and support insights are generated again. The AI insight for each website preview, and the scheduled support-insight audit for each workspace, stopped being written after the move to the new hosting on Oct 17. Sign-in events name how the person actually signed in. A Google, Apple or GitHub sign-in now reads "signed in with Google" (or Apple, GitHub) in the platform events and the Console Activity, instead of "signed in with email" for every account that began with an email address. Password, magic-link and one-time-code sign-ins are named too. GitHub sign-in that matches several accounts explains itself, and you can connect GitHub to your account. When the email addresses on a GitHub (or Google, Apple) account belong to more than one account here, the sign-in page now says so and how to fix it, instead of "We couldn't finish signing you in. Please try again." Settings › Account has a new Connect row to add GitHub, Google or Apple to the account you are signed in to; after that, that provider signs you straight in. Other provider reasons (no verified email, email to confirm, already connected elsewhere) are named too.

ChatShips no longer stall on a brief GitHub hiccup. A deploy now waits and tries again when GitHub returns a temporary error. Before, one such error could stop a release that had already passed every check.

MCPThe connector registry check is fresh again.

DocsYour mate connects your website's own MCP server automatically. When your website publishes an MCP server, Busymate AI now connects it for you, with no setup and no question asked. Your mate then answers visitors from the site's own tools (for example its docs search) instead of crawling its pages. Only tools that really work without signing in are offered to anonymous visitors. The connection appears in Console › MCP servers; delete or disable it there and your site is crawled again.

VoiceBusymate AI widgets catch results that ChatGPT on iPhone sends early, and stop hidden refresh errors. Busymate AI widgets in ChatGPT now keep tool results that arrive before the widget has finished loading, so they show their tables instead of a blank or "Couldn't load" card more often on iPhone. Write results and forms no longer show internal fields such as payload hashes, replay flags or raw IDs.