Anthropic puts programmatic Claude on its own meter — full API rates, June 15.

1. 01

   Anthropic Agent SDK moves to metered credits on June 15

   Full API list prices for any programmatic Claude usage; subscription quotas no longer cover Agent SDK, claude -p, or Claude Code GitHub Actions. Action needed: confirm our billing path before June 15.

2. 02

   Liveblocks Storage v2 migration complete; v3.19.2 fixes a Yjs sync race

   The realtime engine under our canvas finished a full v2 cutover — worth a regression sweep on Storage / Presence / Yjs. v3.19.2 patches a backgroundKeepAliveTimeout Yjs sync drop.

3. 03

   Storybook 10.4: AI-assisted setup, change-detection sidebar, docgen-for-agents

   Storybook is positioning itself as the canonical agent-friendly component truth layer. Strengthens the Onlook + Storybook partnership pitch — and moves toward the visual-review surface we'd want to own.

4. 04

   Cloudflare Workers: Tunnel-in-Wrangler, Domains tab, AI Gateway REST API

   One-key public preview URLs from local dev (t in Wrangler), unified domain management, single endpoint hitting OpenAI / Anthropic / Workers AI with caching and guardrails. Real DX week for our growing CF footprint.

On May 13 (PT), Anthropic announced that programmatic Claude usage — Agent SDK, headless claude -p, Claude Code GitHub Actions, third-party agent harnesses — moves out of subscription quotas (Pro / Max / Team) into a separate, dedicated monthly credit pool, effective June 15, 2026.

Credits inside that pool are billed at full API list prices and don't roll over. Pro ships with $20 of credits, Max 5× with $100, Max 20× with $200, Team Premium with $100/seat. The same week (May 18), Anthropic acquired Stainless — the SDK-generation company powering Anthropic's, OpenAI's, Google's, and Cloudflare's official SDKs — for a reported $300M+, and is winding down hosted Stainless products. Customer-generated SDKs remain owned and modifiable.

Storage v2 — the underlying realtime data engine — is now fully migrated for all rooms (week 20 changelog). Our multiplayer canvas runs on this. The transition wasn't a feature ship, it was an engine swap, so the action this week is verification, not adoption: run our canvas through a real multi-cursor session (sync drops, refresh, large-document edit) and confirm no behavioral drift.

v3.19.2 separately patches a pending-Yjs-update drop under backgroundKeepAliveTimeout; bump if we set that flag.

Two forces in tension. The tailwind: codebases that maintain Storybook are exactly the codebases we want to land in, and Storybook is now actively making itself easier for agents to set up. The headwind: Storybook is moving up the stack toward the visual-review surface — the new change-detection filters are framed explicitly for "agentic workflows where visual correctness can't be automatically validated," which is our use case.

We need to decide whether to consume their docgen + change-detection APIs, integrate with the agent onboarding flow, or compete at the visual-review surface. Worth a Daniel conversation this week.

The Tunnel-in-Wrangler integration — press t in dev to get a public preview URL, with optional Access protection — is the most immediately useful for us. Same-network demos, cross-device QA, sharing a canvas branch with Daniel without staging deploys.

The new AI Gateway REST API (single endpoint hitting OpenAI / Anthropic / Workers AI with caching, logging, rate limiting, guardrails) is worth watching as a unified provider router now that we're consolidating on Anthropic.

v12.39's React 19 drag fix (dragSnapToOrigin no longer strands transforms after layout swaps) and the LazyMotion CJS context fix may quietly resolve canvas-drag jank we may be papering over.

v12.40's new arc() transition is a real capability gain if we ever animate connector lines or node-to-node transitions on the canvas.

Skills are markdown files defining "conventions and workflows you use repeatedly," invoked via slash commands (/insert-sample-data, /build-from-prd) inside any Make prompt. This is the same primitive Claude Code surfaces, now inside the design-to-code surface that competes directly with us. Team / org sharing is "coming soon."

The bigger signal: Figma is moving aggressively into the agentic design-to-code space we've staked. The provisional moat: Skills run against Figma prototypes, not real code — bidirectional code editing on the live React app is still more valuable than design-file synthesis. The open question is whether Figma's distribution and design-file gravity matter more than that premise.

Bolt.new now lets users connect MCP servers inside the editor. That makes MCP-as-extensibility-surface table stakes among visual editors — Bolt, Cursor, Claude Code, Onlook, all running the same protocol. The differentiation has moved from "do you support MCP" to "what does MCP actually buy you inside our surface."

Qwen 3.7 Max, Grok Build 0.1, Chat SDK callback URLs, AI SDK tools subpath, WordPress AI Gateway plugin, Nuxt MCP Toolkit — all shipped in window. Vercel is investing weekly in AI Gateway + Chat SDK + v0 as a wedge against the same designer-developer collaboration loop we own.

We migrated off the Vercel AI SDK, so direct impact is nil — the relevance is that v0 + AI Gateway is the closest first-party platform play in our space.