AI library

Source · release

AI SDK 7

Author
Vercel
Published
25 June 2026
Added
16 October 2026
Original
https://vercel.com/blog/ai-sdk-7

Summary

Vercel's announcement of AI SDK 7, the TypeScript toolkit for AI apps and agents. The release focuses on running agents in production: tool approvals, durable workflows, timeouts, sandboxes, and telemetry. It also adds an experimental way to drive existing agent harnesses such as Claude Code and Codex through one API.

Key points

  • Developing agents: one reasoning option mapped to each provider's native setting; typed per-tool context (contextSchema, toolsContext) so tools see only their own context; typed runtimeContext on ToolLoopAgent; uploadFile and uploadSkill return reusable provider references; MCP Apps rendering in sandboxed iframes (experimental); @ai-sdk/tui for a terminal agent UI.
  • Running agents: per-tool approval policies (toolApproval) with optional HMAC-signed approvals; WorkflowAgent (@ai-sdk/workflow) for resumable runs that survive restarts and delayed approvals; layered timeouts (totalMs, stepMs, chunkMs, toolMs); a portable SandboxSession abstraction.
  • HarnessAgent (experimental): runs harnesses such as Claude Code, Codex, and Pi behind the SDK's Agent interface, with sandbox, skills, and tools configuration and mid-turn interrupt and resume.
  • Observability: registerTelemetry once at startup; @ai-sdk/otel follows the OpenTelemetry GenAI semantic conventions; per-step performance stats such as time-to-first-output and tokens per second.
  • Experimental realtime (OpenAI, Google, xAI) and video generation APIs.
  • Upgrade with npx @ai-sdk/codemod v7 or a migration skill. The post lists no breaking changes; see the migration guide.

Why it matters

AI SDK 7 follows the same direction as the Python frameworks: approvals, durability, sandboxes, and tracing. It also treats complete coding-agent harnesses as components you can embed, rather than building every agent loop from scratch.