Integrating Claude Code with Vercel AI SDK v7

Vercel AI SDK v7 is great if you are building an agent from scratch. But a lot of us already have complex, custom-built harnesses for our agents, systems that handle memory, database persistence, and specialized routing.

The good news is that v7 is highly modular. You are not forced into using their specific WorkflowAgent if you already have a working setup. I tested this by swapping the Vercel AI SDK into a custom harness I built around Anthropic's Claude models.

Dropping in the Core Functions

My old harness manually constructed the Anthropic API payloads and parsed the tool calls out of the response string. It was brittle and broke every time a model changed its JSON formatting slightly.

I ripped out my custom parsing logic and replaced it with generateText from v7.

import { generateText } from 'ai';
import { anthropic } from '@ai-sdk/anthropic';
import { myCustomMemoryStore } from './memory';

class ClaudeHarness {
  async runTask(prompt: string, userId: string) {
    // Fetch conversation history from my custom store
    const history = await myCustomMemoryStore.get(userId);

    const result = await generateText({
      model: anthropic('claude-3-5-sonnet-20240620'),
      system: 'You are a senior codebase architect.',
      messages: [...history, { role: 'user', content: prompt }],
      tools: {
        analyzeFile: {
          /* custom tool logic */
        },
      },
    });

    // Save the new interaction back to my store
    await myCustomMemoryStore.save(userId, {
      prompt,
      response: result.text,
      toolCalls: result.toolCalls,
    });

    return result.text;
  }
}

The Benefit

The integration took about twenty minutes. The SDK handled all the tool parsing and error catching internally, returning clean, typed objects. I got to keep my custom memory database and specialized routing logic without having to maintain the fragile API connection code.

Vercel seems to realize that developers do not want a walled garden. By keeping the core generateText and streamText functions decoupled from their higher-level workflow features, they made it easy to use the SDK as a reliable utility belt inside any architecture.