What’s new

• Names and titles that contain “&” are stored as typed instead of being HTML-escaped by agents.
• The avatar file route is confirmed to require authentication, and that is now pinned by a test.

• The /pr command now sees a pull request that the run opened itself.

• Every build is now type-checked with Next 16 over a tests-free program. No user-facing changes.

The hosted MCP server at /v1/public/mcp now describes its results and errors in machine-readable form and only shows each client the tools it can actually call. AgentWork’s run view and sidebar were reworked to be quieter.

• MCP tools return outputSchema and structuredContent, and errors are structured with typed codes.
• Tool arguments are validated against each tool’s inputSchema before the call runs.
• tools/list is filtered to the tools the caller may call, and its descriptions are slimmed down.
• The plugin’s workflows are served as MCP prompts, and long tool calls send progress notifications.
• launch_task, search and export_run_artifacts have per-tool rate limits; each create_task batch item accepts an idempotencyKey.
• AgentWork: a cleaner run view with quiet tool steps folded into one line, a live per-turn token count, and a rebuilt sidebar.
• AgentWork: chats can start without a folder (each gets a scratch folder), and the macOS app uses a native folder picker.
• Settings opens as a dialog over the current page.

The redesigned interface shipped to everyone and the classic UI was removed, along with the earlier build-time beta and alpha gates that hid some screens.

• Settings screens that were hidden behind a build-time beta gate, including Audio, are now reachable.
• Unused legacy code paths, screens and dependencies were deleted.

A refreshed desktop dock and a fix to the marketplace publish review. Realtime voice remains a preview and is not generally available.

• The dock unfolds from a bare orb, with hover-expand, fact chips and inline approval.
• Marketplace: a gap that let the publish review be bypassed through the MCP tool was closed, and the operator bundle catalog became reviewable data.

• Sign in with Apple, including Hide My Email, can now create an account.
• The /legal/data-deletion page is publicly reachable.
• Admins get a row action to toggle the login-verification exemption for a user.

Crews other developers built are now public at /marketplace — reviewed by us, free to use, and readable in full before you install. Every listing has one sentence to copy: paste it into Claude Code and the crew files land in your project. No account, no plugin, no sign-in; connect AgentTask later and the same crews stay in sync.

• Browse by tag or search; every card shows the personas, the creator and how many workspaces installed it.
• One copyable sentence per listing, a Cowork variant, and a “do it by hand” file list.
• /bundles now redirects here; paid AgentTask bundles stay in the app’s Crews page (the Marketplace tab).

Ready-made packs of expert AI crews are now easy to find. The Crews page has a Store tab next to your own crews, the New crew menu can jump straight to it, and each bundle opens a details panel that shows exactly which personas are in the box before you buy. The catalog is also published here on the site.

• Crews → Store: browse and buy bundles, deep-linkable with ?tab=store.
• Bundle details show every persona’s role, trigger and mode — never its instructions.
• The public catalog moved to /marketplace on 2026-09-19; /bundles redirects there.

• Dependency and tooling updates, and a re-pin of the desktop MCP engines. No user-facing feature changes.

The desktop AgentWork engine got a large round of capability and safety work: a richer conversation view, per-tool permissions, an integrated terminal and browser, and tighter control for organizations.

• A transcript thread with live composer, steerable queue and per-message actions.
• Per-tool permissions with rule promotion and a review inbox, plus an effective-permission view that shows which layer decided.
• MCP dynamic registration, hosted MCP Apps UI and per-workspace authorized folders.
• An integrated terminal dock, an embedded browser panel, retained session tabs and a side-by-side split.
• Organization policy: an org-managed web origin allow-list, org-authored blocked-command patterns, and retention ceilings.
• Crash recovery, a hardened updater and signed connect links that bind an install to an organization.

Crews are assignable AI teammates: define a persona once — instructions, deliverables, when it gets involved — link its skills as notes, and assign it to tickets like a person. Claude Code picks the persona up over MCP, does the work in advise or execute mode, and reports back on the ticket with clear AI attribution. /crews sync even materializes your crews as native Claude Code subagents with enforced tool permissions.

• Crews live in your space sidebar with codes like AI-CR1.
• Advise-mode crews start from read-only tools (Read, Grep, Glob); any extra tool is an explicit grant.
• The human assignee stays responsible; the crew works alongside.

The site you are reading went all-light: warm paper background, iris and teal accents, a live agent-run window in the hero, and new pages — Solutions, Migrate, Customers — plus a crew of cartoon robot mascots who now handle our 404s.

• Homepage rebuilt to a tighter 9-section structure.
• Pricing published with the new packaging.
• Official product logo everywhere.

Notes became a first-class knowledge layer: opt-in version history with restore and diff, threaded comments with reactions, attachments, an activity log, and universal links that connect notes, tasks, and projects in both directions.

• Built and hardened across two rounds of work.
• Versioning snapshots are atomic; restores are one click.
• Agents read and write notes over MCP with the same access rules as people.

Projects gained threaded comments with reactions, an activity feed and audit log, inline title editing, and a timeline view that lays your initiatives out on a real calendar.

• Comment on a project like a ticket.
• Timeline math handles overlapping initiatives.
• Everything is audited, like all agent-adjacent surfaces.

• New physics-based task visualization: tasks float as nodes in organic group membranes, with status colors, drift, zoom, and a magnifier.
• Drag a group to move it; gravity re-settles the cluster.

• Browsable API reference and OpenAPI document.
• Remote MCP endpoint for agents and clients.

• Organizations, spaces, members, projects, and groups.
• Agent claims, heartbeats, and audit trail.