wigolo as Vercel AI SDK tools — register webSearch / webFetch / research on any model

wigolo-vercel-ai-sdk wraps wigolo as ready-to-register Vercel AI SDK tools. WigoloMcpClient spawns and talks to a wigolo instance over the MCP protocol on stdio (npx wigolo by default — no server to run), and createWigoloTools hands you the tool set any AI SDK model can call.

Run it

npm install
npm run demo     # tsc && node dist/tools.js

Requires node >= 20. No LLM key needed for this demo — it connects, registers the tools, prints them, and disconnects.

What you'll see (real output)

connected to wigolo over MCP (stdio)

registered 6 tools:

- webSearch(query, max_results, include_domains, exclude_domains, ...)
    Search the web for information on any topic.
- webFetch(url, section, render_js, max_chars, ...)
    Fetch a specific web page and return clean markdown content.
- webCrawl(url, strategy, max_depth, max_pages, ...)
    Crawl a website starting from a URL.
- findSimilar(url, text, max_results)
    Find pages semantically similar to a given URL or text from the local cache.
- research(topic, max_depth, max_sources)
    Deep multi-step research on a topic.
- agent(goal, max_steps)
    Autonomous web agent that breaks down complex goals into search/fetch/extract steps.

disconnected — wigolo subprocess stopped

Using the tools with a model

With any AI SDK provider wired up, giving the model live web access is one property on generateText / streamText:

import { generateText } from 'ai';
import { anthropic } from '@ai-sdk/anthropic'; // or any AI SDK provider
import { WigoloMcpClient, createWigoloTools } from 'wigolo-vercel-ai-sdk';

const client = new WigoloMcpClient();
await client.connect();

const { text } = await generateText({
  model: anthropic('claude-sonnet-4-5'),
  tools: createWigoloTools(client),
  maxSteps: 5,
  prompt: 'What changed in the latest TypeScript release? Cite sources.',
});

await client.disconnect();

The model decides when to call webSearch / webFetch / research; wigolo does the gathering locally — ML-reranked search, clean markdown extraction, a persistent knowledge cache — and returns structured, citable results the model can quote. You can also cherry-pick single tools (createWebSearchTool(client) et al.) instead of registering all six.

Notes

The subprocess is plain stdio MCP — the same wigolo that Claude Code, Cursor and friends register directly. This package just adapts it to the AI SDK's tool() interface with zod-typed parameters.
Point the client at a specific wigolo entry with new WigoloMcpClient({ command, args }) (the demo reads WIGOLO_MCP_COMMAND / WIGOLO_MCP_ARGS for this).
Prefer HTTP instead of a subprocess? Use rest-curl or the typed SDKs: sdk-typescript-research.

Vercel AI SDK tools

Run it

What you'll see (real output)

Using the tools with a model

Notes

On this page