Vercel AI SDK 7 Ships Native MCP Support — Plus Anthropic Rate Limit Parity and Microsoft's New Copilot Model

TL;DR: Three dev-AI platform updates: Vercel AI SDK 7 adds native MCP and durable agents; Anthropic equalized rate limits across all Claude tiers; Microsoft launched its first proprietary coding model for Copilot Business/Enterprise.

Vercel AI SDK 7 — native MCP, WorkflowAgent

Vercel AI SDK 7 (June 25) is the production TypeScript/React layer for building AI apps, with 16M+ weekly downloads. Version 7 adds:

• Native MCP support — connect any MCP server with experimental_createMCPClient(); tool-level permission control per server
• WorkflowAgent — durable agent workflows that persist state across steps and serverless function boundaries
• Tool approvals — middleware to intercept and gate tool calls before execution, enabling human-in-the-loop patterns
• Provider files — centralize model routing and fallback logic in a config file instead of scattering it across code

Same useChat/useCompletion API whether you're targeting Claude, GPT-4o, or Gemini.

Why it matters: MCP lands in the most widely used TypeScript AI SDK. Web apps built on AI SDK 7 can now directly consume MCP servers — including ones you write.

Anthropic API rate limit parity

As of June 26, Claude Sonnet and Haiku rate limits now match Claude Opus at every usage tier. Sonnet 4.6 has been the price/performance favorite for most agent tasks; hitting its tighter rate ceiling (relative to Opus) was a common friction point for teams running high-throughput agent pipelines. That asymmetry is gone.

Why it matters: You can route more agent traffic through Sonnet without hitting a rate cap that Opus wouldn't have.

MAI-Code-1-Flash for GitHub Copilot Business/Enterprise

Microsoft AI's purpose-built coding model is now GA for Copilot Business and Enterprise (June 26). MAI-Code-1-Flash is optimized for low-latency, high-volume iterative agentic workflows — the pattern where a coding agent calls the model thousands of times per task. Org admins enable it via Copilot settings; billed on usage.

Why it matters: Benchmark it on your actual tasks before switching — low latency is real, but capability tradeoffs exist for heavier reasoning.

Sources: Vercel AI SDK 7 | MAI-Code-1-Flash — GitHub Blog | Anthropic platform updates — Releasebot