Vercel releases AI SDK 7 with new tools for managing AI agents

Vercel has released version 7 of the AI SDK, its open-source TypeScript toolkit for building AI applications and agents. The SDK sees more than 16 million downloads a week and also powers eve, Vercel's open-source agent framework. The release adds reasoning control, support for MCP Apps, and a push beyond text-only agents into voice and video.

What's new for building agents

AI SDK 7 standardizes reasoning control. Most top models support configurable reasoning, but every provider exposes it through a different API. Now the reasoning option in the generateText and streamText functions takes a single line of code (for example, reasoning: 'high') and maps automatically to each provider's native settings. Developers who need finer control can still set provider-specific parameters manually.

• More than 16 million AI SDK downloads per week
• A single reasoning option replaces each model's provider-specific settings
• Typed toolsContext: data such as API keys is passed to a specific tool via contextSchema, without entering the model's prompt or leaking to other tools
• Upgrade with pnpm add ai@latest; migrate from version 6 with npx @ai-sdk/codemod v7 or the skill npx skills add vercel/ai --skill migrate-ai-sdk-v6-to-v7

What's new for running and observing agents

Beyond development, AI SDK 7 expands what's available for running agents in production: approval for tool calls, fault tolerance via WorkflowAgent, timeouts, and support for sandbox environments. The SDK can also integrate with any third-party agent harness, including Codex, Claude Code, Deep Agents, OpenCode, and Pi.

To monitor agent behavior, the release adds telemetry, a Node.js tracing channel, lifecycle events, and performance statistics. Finally, AI SDK 7 moves beyond text-only agents with provider-agnostic support for real-time voice agents and video generation.

What it means

AI SDK 7 shows where agent tooling is heading: away from simple prompt-based prototypes and toward production components with reasoning control, auditable tool access, fault tolerance, and full telemetry. Without these, an agent can't safely be put in front of real users.

FAQ

How do I upgrade from AI SDK 6 to AI SDK 7?

You can run the automated migration with npx @ai-sdk/codemod v7, which rewrites your code with minimal changes, or use the skill: npx skills add vercel/ai --skill migrate-ai-sdk-v6-to-v7.

What is toolsContext in AI SDK 7?

It's a typed context defined separately for each tool via contextSchema. For example, an API key is passed directly to the tool's exec function rather than through the model's prompt, and other tools can't see that context.

Which agent frameworks does AI SDK 7 support?

The SDK can integrate with any agent harness, including Codex, Claude Code, Deep Agents, OpenCode, and Pi.