Skip to content

The AI Platform 1.0: specialists that remember, a richer browser, and a PlanetScale suite

Specialists start to remember across conversations, gain a much deeper browser toolset, and a PlanetScale specialist suite joins the marketplace.

This is The AI Platform 1.0. Specialists are starting to preview memory that carries across conversations, the in-app browser toolset gets a major expansion, and a PlanetScale specialist suite joins the marketplace. Workflows, billing, knowledge, and feedback all pick up sizable upgrades on top.

Specialists are starting to remember

We are starting to preview memory for specialists in a small set of eligible sessions. When it is turned on for you, specialists can save useful context as you work and recall it on later turns, so you do not have to keep restating who you are or what you are building. Memory is gated by role, auto-saves only what is worth keeping, and redacts secrets and personal data on the way in. We will widen the rollout as we get more usage in.

A PlanetScale specialist suite in the marketplace

A new PlanetScale specialist suite is available in the marketplace, with a parent specialist and dedicated Postgres and Vitess children. Add All installs the full set in one click. The parent delegates focused work to the right child, and the child’s progress, the resources it touches, and the tools it uses all render in line under the parent’s turn so you can follow along. The suite ships with PlanetScale documentation as a knowledge source, with clear queued, indexing, indexed, and failed states as it comes online.

Spawned specialists are now scoped to the workspace and conversation that started them, declined immediate-approval requests resolve cleanly as cancellations, and changing crawl depth on a knowledge source resets and requeues so you get a clean re-index.

Specialists, sharper across the board

Long specialist instructions now condense gracefully, so specialists with very long personas stay responsive instead of getting truncated mid-turn. You can invite channel members by @mention straight from the composer, with tighter checks on who is allowed to be added. Every specialist is visible in private channels (some were previously hidden), and the bundled welcome specialist now survives app updates without losing its slug.

A much deeper browser for specialists

Specialists can now do a lot more inside the in-app browser. They can navigate pages, inspect the DOM and computed styles, read the console, watch network requests and responses, read and write cookies and storage, emulate a different device or location, capture screenshots with viewport, clip, format, and quality controls, and run accessibility and performance queries. Every operation is gated by its own permission, so you stay in control of what a specialist can see and do.

Tool use is also more reliable: a registration bug is fixed, sites that previously failed to load now work, and the approval prompt shows a real tool name and a message filled in with the actual arguments — so you know exactly what is being requested before you say yes. App panels and the embedded view also stay alive when you close them, so reopening is instant and you do not lose browser state.

Workflows: structured outputs and saved artifacts

The Specialist Invoke node now has first-class structured output. Downstream nodes can reference json and structured fields directly, the Knowledge Garden node supports a passthrough output mode, and there is a clearer choice of invocation modes for how a specialist is called. Large node outputs are now stored as artifacts and surfaced in Run History, so debugging a long workflow no longer means scrolling through a wall of text.

Knowledge Garden: URLs and repositories as sources

You can now point Knowledge Garden at a URL or a code repository as a first-class source root. Crawled pages appear as children of the URL, repository files appear with their repo-relative paths and current git status, and preview, download, delete, and refresh all work on source-backed items without disturbing unrelated uploads. The “plot” action is also reachable from more surfaces, not just the sidebar.

Chat polish that adds up

If a streaming reply is interrupted, the partial content is preserved so you do not lose what was already written. Attribution notifications no longer ping you while you are already looking at the conversation. And the provider that answered a given message is remembered across reloads, so the badge does not change shape on you.

Sandbox and uploads

A new chat-level sandbox control lets you override execution sandbox behavior per chat, and there are safer prompt and environment overrides for QA-style runs. Workspace file uploads now route through the local app, which makes them more reliable on flaky networks. Workspaces can share a display name (the uniqueness requirement was unnecessary), and you can upload a custom image as a project icon. Your account avatar now propagates everywhere you appear: member lists, direct messages, and mentions.

Providers and models

Secondary provider groups in provider configuration collapse by default, so the list of providers is much easier to scan. The canonical model catalog has been refreshed. OpenRouter PDF attachments and the file-parser plugin no longer drop out of the pipeline (a regression where PDFs were being misclassified as a generic file type has been fixed).

Billing, refreshed

Plan management has been overhauled. The credit usage bar is wider and easier to read, the plan selector page is cleaner, and there is a pending state while a new subscription is being created so you are never left wondering. Coupons are supported when changing plans, with clearer “due now” dialogs and an invoice preview for the next cycle.

You can now create subscription offer links that pre-fill the plan, billing cycle, seat count, and coupon code — useful for handing prospective customers (or enterprise contracts) a one-click signup. Promotion emails have simpler titles, customers without a payment method see a helpful tooltip, and a failed payment now shows a clear retry warning with up to four retries before cancellation. Workspaces that have billed in the past keep access to billing surfaces even without an active subscription, and subscription-only options are no longer shown to trial or free workspaces. Billing also no longer hangs on a slow startup step.

Auth and security

Shared provider credentials now sync correctly across your devices, with hardened authorization on the backend. If your account is auto-joined to a workspace because of your email domain, you get a notification about it. The OAuth sign-in success page no longer flashes a broken theme on its way in. And as a defense-in-depth measure, your authentication token is no longer exposed to the frontend layer.

Feedback reports with optional logs

When you file a feedback report, you can now attach the last 5, 15, or 30 minutes of app logs with one click. Sensitive values are redacted before anything leaves your machine. No more zipping log files by hand.

Polish & fixes

If you are still running the Intel build on an Apple Silicon Mac, the app now prompts you to switch to the native build before macOS stops running it. The app no longer flashes when a dialog opens over the embedded browser. Long workspace names on the homepage greeting now break onto two lines instead of getting cut off. New workspace seeding uses the same path as the rest of the app, so a fresh workspace is consistent from day one. Project forms get a pass on inputs, button sizes, hover, padding, and gaps so they feel unified. The offer page picks up a visual refresh, and the pricing page is redesigned with accessibility and performance improvements.