What shipped.

LLM steps emit structured AI generation telemetry (including PostHog) across chat, agents, and artifacts so production evals and cost reviews share one signal path.

Pull a skill pack from GitHub straight into the agent file tree, discover/load skills at runtime, and pick skills from the composer slash menu.

Skills v0.2.0

Mobile agent-first shell, public storage/deliverable shares with deep links, Storage explorer/preview polish, and Pumble /manifest marketplace install land together for the evals-to-prod release.

POST …/agents/runs/:runId/interrupt stops an in-flight turn cleanly so clients can cancel or steer without orphaning the stream.

Managed Agents API v0.3.0

Storage groups agent-written files with run artifacts under Deliverables instead of duplicating them in Files.

Chat forwarding catches upstream failures and returns JSON with a useful cause, so failed agent turns fail loudly instead of silently.

Shared workspace links open the right file or session via ?fileType=&fileId= and ?session=; public previews can send members back into the workspace or Ask an agent.

Sessions, deliverables, and reports always show a workspace Copy link after auth, with harder share APIs and table fallbacks when public settings fail.

Workspace agents v0.5.1

generate_api_key now qualifies extensions.gen_random_bytes so creating workspace API keys works again in production.

Equal-width selector chips, larger touch targets, viewport-capped popovers, and a drawer with workspace switcher - desktop layout stays unchanged.

Workspace agents v0.5.0

Dashboard gets KPI meters, a compact apps list, a unified activity feed, and clearer widget settings.

Find threads by title and export conversation data more easily from Workspace Insights.

Reports v0.4.0

Skills and tools share clearer browse/installed panels so attaching capabilities to a file agent is consistent end to end.

Skills v0.1.0

APIs under /agents/registry and /agents/seed-default let you list file agents and backfill the core agent on existing workspaces.

Managed Agents API v0.2.0

Mention folders from the tree, review file-change cards in the timeline, and run plan review from the plan markdown toolbar after a research phase - composer modes are Auto, Plan, and Ask.

Workspace agents v0.3.0

Connect flows accept returnTo and the callback redirects safely back to the requested agent/config path.

MCP v0.2.1

OAuth redirect handling is tighter for sign-in and integration install flows, reducing land-on-the-wrong-page failures.

Pumble integration ships with production-ready install/setup seeding and marketplace branding so workspaces can connect the app without a one-off ops path.

MCP v0.2.0

Leaderboard views get clearer charts and display polish so model comparisons are easier to scan.

Reports v0.3.1

Connect agents to MCP-connected tools through a managed integration connection, instead of one-off per-agent config.

MCP v0.1.0

Refactored authentication handling and middleware for more consistent behavior across routes.

Start a new agent session from a template, with more robust session state handling underneath.

Workspace agents v0.2.0

Upload images and documents from the composer or via API - png/jpeg/webp/gif and txt/md/csv/json/pdf up to 15MB, backed by signed URLs.

Managed Agents API v0.1.0

The agent panel now defaults open and stays in sync on expand/collapse, instead of intermittently losing chat history.

Workspace agents v0.1.1

Client-side caching and SQL-side aggregation cut redundant queries; pages render from cache instantly and revalidate in the background.

Agent runs using newer gateway model specs (e.g. Claude Opus 4.6 via the AI Gateway) no longer crash on legacy streaming paths.

Multi-model routing v0.0.2

Traces now capture time-to-first-token, action outcomes, and normalized error codes, with an error taxonomy and transient retries.

Tool rows stay collapsed by default during streaming and the timeline reads denser, with less visual noise between steps.

Workspace agents v0.0.2

Workspace invites and notifications now send branded emails, with an acceptance flow and specialized templates per notification type.

A purpose-built mobile experience for Agents - chat, history, header, input bar, and sidebar tuned for phones.

Workspace agents v0.0.1

Agents can now connect to external systems via Model Context Protocol connectors.

MCP v0.0.1

A complete design system bundle with previews for key pages and components, laying the groundwork for the rest of the redesign.

Agents can attach reusable skill modules, sourced from the workspace or directly from a GitHub repo.

Skills v0.0.1

POST /api/v1/artifacts/generate turns customer aggregates and samples into a structured executive narrative report.

Reports v0.3.0

When the primary gateway is not configured, the platform now auto-selects Gemini or Groq based on the model id and available keys.

Multi-model routing v0.0.1

New agent APIs under /api/v1/workspaces/:workspaceId/agents for chat, session management, and run lifecycle (events, cancel).

Managed Agents API v0.0.1

Wallets can auto-recharge automatically, with enhanced balance and usage endpoints in the billing UI.

Credits v0.1.0

Save a report once and regenerate it on demand from the editor preview, instead of rebuilding it from scratch each time.

Reports v0.2.0

Embedded widgets now stream through the same v1 chat pipeline as the rest of the platform, with external identity and page context support.

Assistant API v0.2.0

Workspace credit balances and ledger with automatic per-request debiting, low-balance handling, auto-recharge, and manual top-ups.

Credits v0.0.1

Standardized token pricing slugs across model ids so models like openai/gpt-oss-120b resolve pricing correctly instead of billing misses.

Credits v0.0.2

Review threads across apps in one place, with filters, search, and a rich conversation preview (messages, metadata, timeline).

Reports v0.1.0

Per-app ScaleDown settings reduce token usage and cost automatically, with savings and latency shown in overview analytics.

Reports v0.0.1

A dedicated /workspace/:id/api/image page alongside the Text API, sharing the same subnav and key management.

Images API v0.0.1

A dedicated /workspace/:id/api page with its own subnav, consolidating API key management for chat/text.

Assistant API v0.1.0

Public-path checks, improved CORS headers, and smarter post-login redirects across the workspace-first flow.

Moved to the shadcn radix-vega design system with OKLCH color tokens, refreshed typography (Inter/Lora via next/font), and hardened OAuth flow.

Public REST endpoints under /api/v1/apps/[appId]/... for chat (with SSE streaming), memory threads/messages, and vector index search, with shared API-key auth.

Assistant API v0.0.1

Polish pass on assistant chat chrome based on widget and editor feedback.

Assistant API v0.0.0-beta.2

Widget hardened with streaming responses and external identity support before the v1 bridge at GA.

Assistant API v0.0.0-beta.2

Visual and interaction updates to the embeddable widget shell.

Assistant API v0.0.0-beta.2

Callable image generation API - precursor to the dedicated Workspace API Image tab at GA.

Images API v0.0.0-rc.1

Named release-candidate train consolidating editor, chat, and billing for GA readiness.

Refreshed vertical case studies and landing pages with updated proof points.

Maintenance release: dependency updates and stability fixes during a lighter shipping month.

Schema migrations laying groundwork for upcoming product surfaces and API expansion.

Documents exportable from the editor for offline sharing and deliverables.

Refreshed marketing pages to reflect pilot and insights capabilities.

Generate PDF exports from editor content for customer deliverables.

Export layout and formatting refinements based on pilot feedback.

Fine-grained permissions for pilot participants across workspace surfaces.

Pilot cohorts gated by role so design partners see only the surfaces they are approved for.

Centralized Supabase client usage and cleaner navigation components for maintainability at pilot scale.

Published case study pages on the marketing site with vertical-specific proof points.

Enrollment UX and routing so interested teams move from marketing page to an active pilot cohort.

ROI calculator surfaced in primary navbar and footer navigation.

Metering matured with Stripe usage records ahead of the credit wallet GA.

Credits v0.0.0-beta.1

Workflows capability linked from marketing chrome for broader product discovery.

Visual and interaction refinements to the editor chrome alongside chat beta usage.

Assistant API v0.0.0-beta.2

Reduced time-to-interactive when opening a workspace after login.

Updated marketing landing layout and content for the chat-and-collab beta push.

Workspace navigation adapts to mobile with a sheet-style sidebar.

Multiple authors can edit the same document concurrently with live collaboration sync.

Jump back to recent documents without hunting through the full file tree.

Authors can resize the chat panel within the editor layout.

Assistant API v0.0.0-beta.2

Interactive ROI tooling on the marketing site to frame value during pilot scoping.

Documented local Supabase workflow so contributors and pilot engineers can run the stack offline.

Swapped default model references from Mixtral to Llama 3.3 70B as provider quality shifted.

Additional formatting and writer tooling in the editor beyond the initial MVP.

Additional formatting and writer tooling in the editor beyond the initial MVP.

Panel layout controls so authors can tune the editor workspace on different screens.

Fast Groq inference wired through the AI SDK for responsive in-editor generation.

Marketing copy and structure aligned with the editor-first product direction.

Core authoring surface with the file structure and UI patterns later chat and APIs extended.

Commercial checkout and subscription plumbing so pilots could convert to paid without a billing rewrite.

Workspaces isolate customer data with role-based permissions from the first release.