How do I give a Vercel AI SDK app web access with UpRock?

What the model gets

UpRock's MCP server runs over streamable HTTP at https://mcp.uprock.ai/mcp, with your API key in an Authorization: Bearer header (clients that can't send headers can use https://mcp.uprock.ai/<API_KEY>/mcp instead). client.tools() returns an object you can pass straight to generateText, with these keys:

To expose only some tools to the model, pick them out of the object, for example { crawl_fetch: tools.crawl_fetch, resource_fetch: tools.resource_fetch }.

The code

The example makes a direct call first (no model involved), then lets Claude choose the tools in generateText. Set UPROCK_API_KEY (from uprock.ai/app/api-keys) and ANTHROPIC_API_KEY, or swap in any other provider:

// npm i ai @ai-sdk/mcp @ai-sdk/anthropic   (Node 18+)   env: UPROCK_API_KEY, ANTHROPIC_API_KEYimport { createMCPClient } from "@ai-sdk/mcp";import { anthropic } from "@ai-sdk/anthropic";import { generateText, isStepCount } from "ai";​const uprock = await createMCPClient({  transport: {    type: "http",    url: "https://mcp.uprock.ai/mcp",    headers: { Authorization: `Bearer ${process.env.UPROCK_API_KEY}` },  },});​type Block = { type: string; text?: string };const textOf = (result: unknown) =>  (result as { content: Block[] }).content.find((b) => b.type === "text")?.text ?? "";​try {  const tools = await uprock.tools(); // crawl_fetch, resource_fetch, video_search, ...​  // 1. A direct call, no model involved: load a page on a device in Germany. Tool calls have  // no default timeout, so pass one.  const options = {    toolCallId: "1",    messages: [],    context: {},    abortSignal: AbortSignal.timeout(300_000),  };  const page = JSON.parse(    textOf(      await tools.crawl_fetch.execute!(        { url: "https://example.com", country: "DE" },        options,      ),    ),  );  const markdown = textOf(    await tools.resource_fetch.execute!(      { uri: page.content.markdown.resource },      options,    ),  );  console.log(page.meta.title, markdown.length, "chars of markdown");​  // 2. The same tools, chosen by a model.  const { text } = await generateText({    model: anthropic("claude-sonnet-5-5"),    tools,    stopWhen: isStepCount(5),    system:      "When you need a web page as it looks in a specific country, or another fetch failed, call " +      "crawl_fetch with the url and country, then resource_fetch on the markdown resource it returns.",    prompt: "What does https://example.com say when opened from Germany?",  });  console.log(text);} finally {  await uprock.close(); // after generateText resolves; with streamText, close in onFinish}

Calling execute yourself returns the raw MCP result, { content, isError }. In the AI SDK 7 types, its options need toolCallId, messages and context; abortSignal is optional but worth passing.

What comes back

The first text block of a crawl_fetch result is a JSON summary, small on purpose:

{  "status": "success",  "meta": { "url": "https://example.com", "title": "Example Domain", "content_type": "text/html" },  "content": { "markdown": { "resource": "crawl://…/markdown", "size_bytes": 1342 } }}

When the model calls the tool, the AI SDK turns text into text and everything else, including resource links, into JSON text, so the model sees this summary and the resource address. The system prompt in the example tells it to follow up with resource_fetch, which returns the page as markdown. Don't rely on meta.status_code: it isn't always present.

Good to know

• No default timeout. Pass an abortSignal on direct calls; for model-driven calls, pass one to generateText.
• Close the client after generateText resolves, or in onFinish with streamText.
• The key is in a header. Keep it in an environment variable, and don't log the client's config or MCPClientError objects whole.
• Cost. One credit per page fetched: $0.50 per 1,000 pages by card or USDC, $0.40 with UPT. Cap the agent's steps (isStepCount) and say in the system prompt when to ask before fetching many pages.

We ran this example on 21 Oct 2026 with ai 7.0.148 and @ai-sdk/mcp 2.0.81 against the live UpRock server. The direct calls returned example.com's title and markdown. For the generateText part we used the AI SDK's own mock model, scripted to call crawl_fetch and then answer, in place of Claude: the loop ran two steps and the tool call went through UpRock. On 6 Nov 2026 we ran the direct calls again with an UpRock API key in an Authorization header.