Vercel AI Updates: May 27, 2026

1. Vercel Sandbox persistence goes GA, defaulting to durable filesystem state with named resume and fork APIs

Vercel. Vercel Sandboxes now preserve filesystem state across sessions by default, with persistence enabled out of the box. Developers can create, retrieve, or resume sandboxes by durable name, and operations like runCommand() or writeFiles() on stopped sandboxes auto-resume from the most recent snapshot. New APIs include Sandbox.fork(), Sandbox.getOrCreate(), and Sandbox.delete(), plus onCreate and onResume lifecycle hooks and custom tags for multi-tenant tracking. Snapshot storage is billed separately, and ephemeral workloads can opt out via persistent: false or the --non-persistent CLI flag. Source

2. Firecrawl joins the Vercel Marketplace as a first-party AI web-scraping integration

Vercel. Firecrawl is now available as a first-party integration in the Vercel Marketplace, giving AI agent and application developers structured web data without standing up crawling infrastructure. The service converts pages into markdown, HTML, structured data, or screenshots, and can run a web search plus full-page retrieval in a single operation. It also supports interacting with dynamic sites through AI-driven prompts or custom code, targeting retrieval and agent workflows that need fresh, LLM-ready content. Source

3. AI SDK MCPClientError gains statusCode, url, and responseBody for structured HTTP failure context

Vercel. The @ai-sdk/mcp 2.0.0-canary.55 release adds three optional fields to MCPClientError — statusCode, url, and responseBody — when errors originate from the streamable HTTP transport. The change lets downstream agent frameworks read the actual HTTP response status without string-parsing error messages, improving diagnostics for remote MCP server failures. The new fields remain undefined for stdio transport errors and non-response failures like network errors or aborts, preserving backward compatibility. The release shipped on the canary track alongside coordinated bumps to ai@7.0.0-canary.154 and related framework adapters. Source

4. AI SDK OpenTelemetry helper now wraps model calls in the active trace context

Vercel. The @ai-sdk/otel 1.0.0-canary.100 release updates the integration so model calls are wrapped in an OpenTelemetry context, ensuring spans emitted by AI SDK operations are properly parented under the caller’s active trace. The change improves correlation between application-level traces and the underlying LLM request spans in OTLP-consuming observability tools. The release shipped on the same canary train as ai@7.0.0-canary.154 and related @ai-sdk/* adapters, indicating a coordinated SDK version bump across the workspace. Source