The AI Platform 2.0: rebuilt messaging, notifications that respect your attention, and a platform to build on

The AI Platform 2.0 is a major release. The biggest change is one you mostly feel rather than see: messaging has been rebuilt on a new backend that is now ready to run everything you do in channels, threads, and direct messages — so conversations load faster, stay in sync across your devices, and hold up when a busy room or a whole cohort piles on at once. On top of that foundation, notifications now respect your attention, and the platform you build on — miniapps, the in-app browser, and workflows — took a big step forward.

Messaging is faster and more reliable

Channels, threads, and direct messages now run on the rebuilt messaging backend, and the practical result is fewer stalls and less drift. Direct messages deliver in real time again, dropped connections reconnect calmly, and conversations no longer hang waiting on a single slow response. Opening a conversation is quicker too: cached messages show instantly while fresh ones arrive in the background, and editing or deleting a message updates the timeline right away.

Notifications that respect your attention

Notifications got a ground-up pass so they point you to the right place and stop counting things you have already handled. Native banners and sounds are back for mentions and direct messages, a thread mention takes you straight to that reply, and opening a conversation now clears all of its notifications from the bell and app badge. Your mute, mentions-only, and Do Not Disturb preferences are finally honored everywhere, and a new Notification types checklist lets you switch off whole categories you do not care about.

A smoother composer and cleaner transcript

The composer and transcript are calmer to write and read in. Mention and shortcode popovers follow your cursor in long drafts, @-mentioning a file now searches your whole project instead of just the current folder, and a screenshot you paste stays put if you leave the conversation before sending. Messages themselves read better: compact grouping is back, spacing and hover states are consistent, and specialist replies again show their model, provider, token usage, and cost.

Channels, threads, and direct messages

Channel management now behaves the way you would expect from a mature workspace. Archiving and restoring channels is admin-owned — regular members use Leave Channel — and the full Channel Details dialog is reachable from the channel menu again. Adding someone to a channel now puts it in their sidebar right away, and channel membership stays consistent everywhere it is shown. You can create channels inside a sidebar group and drag them between groups, and leaving a private channel asks you to confirm first.

Specialists and Chloe

Specialists hold onto more context and surface more sensibly. A specialist no longer forgets your configured project folder after a restart, the old assistant/operator/orchestrator labels are gone in favor of what a specialist can actually do, and specialists draw on a real memory now — ranked by a blend of relevance, recency, and retention rather than plain keyword matching. Chloe, the built-in assistant, keeps responding even on a fresh workspace with no plan selected: it runs on a dedicated allowance with sensible daily and monthly limits and falls back to other models when needed.

Build on the platform: miniapps and workflows

If you build on The AI Platform, 2.0 is a big one. A new miniapp platform underpins first-party and third-party apps, with an install and marketplace flow, per-app permissions, host-owned storage, governed network and credential access, and a supported UI component library so your app matches the platform out of the box. Miniapp settings are now split into Installed, Discover (backed by a verified, remotely managed catalog), and Custom (including local ZIP drag-and-drop).

For builders, miniapps can now make network requests to services you have approved and use stored credentials without ever receiving the secret values, ship tools that your specialists call during a chat, and be built in Rust and WebAssembly against the shared component library. Two first-party apps ship on top of the platform: API Workbench, a REST and GraphQL client with collections, environments, and a collection runner that specialists can drive, and a Design File Viewer for inspecting local design files.

Workflows are steadier as well: disconnected action nodes are skipped instead of failing a run, workflow names are validated, and you can cancel a run and re-run from history with clear feedback.

The in-app browser and Canvas

The in-app browser grew into a real workspace tool, with browser profiles you can assign per project or channel, one-click import of your existing cookies, and channel collaboration where one person or specialist holds control at a time. Canvas adds collaborative drawing, versioned UI mock iteration, and review with history and comments.

Providers and models

When a model is overloaded or a provider fails, you now get a way forward instead of a dead end. If every route for your selected model is exhausted, The AI Platform ranks alternatives from your personal, workspace, and specialist preferences and offers a one-click retry on a recommended model. Your account picture is now consistent everywhere it appears, and the model catalog was refreshed.

Sign-in and account

Signing in is more durable and faster. The desktop app stays signed in across restarts, and if the secure credential store and account record ever disagree, the app repairs itself instead of silently dropping you to the login screen. Everyday actions no longer stall while your session refreshes in the background, and cold sign-ins are noticeably quicker.

Performance and stability

This release folds in a broad round of performance and stability work. File operations behind specialists and tools are measurably faster, and file writes and edits no longer time out. Clicking a download button no longer freezes the app, a macOS crash when opening a document attachment is fixed, and on Windows a repository no longer flashes repeated console windows and gets stuck until restart. Crash and error reporting now reaches us reliably, and billing usage is recorded more reliably under heavy load.

Polish & fixes

Button heights and shadows are consistent again, and the app icon is re-centered in the macOS tile. The Tasks page dropped a duplicate filter control, and newly created tasks save reliably. Opening a shareable workspace invite link now joins you immediately. A silent corruption of shared knowledge graphs on import is fixed, migrated channels no longer list a system account as a member, and member names resolve to display names rather than raw emails.