Models

Which provider packages?

Use the Vercel AI SDK’s own provider packages. typemaid does not ship providers and does not need to: typemaid/ai is a thin wrapper that accepts any AI SDK model.

npm install ai @ai-sdk/openai
import { openai } from '@ai-sdk/openai';
import { defineAI } from 'typemaid/ai';

export const ai = defineAI({ model: openai('gpt-5-mini') });

To switch provider, install another package (@ai-sdk/anthropic, @ai-sdk/google, …) and change that one line. New models work as soon as the provider package supports them, with no typemaid release.

ai is an optional peer dependency. Workflows that never call a model do not need it, and it is only loaded when the first model call runs.

An agent step

A step whose work is one model call with a typed answer:

export const summarize = ai.agent({
  prompt: 'Summarize in three bullets: {text}',
  output: z.object({ bullets: z.array(z.string()) }),
});

{text} is filled from the shared data. The checker reports a placeholder nothing wrote.

A model call inside a tool

export const reply = tool({
  input: z.object({ review: z.string() }),
  output: z.object({ reply: z.string() }),
  run: async ({ review }, ctx) => {
    const { sentiment } = await ai.ask(ctx, {
      prompt: `Is this review positive or negative? "${review}"`,
      output: z.object({ sentiment: z.enum(['positive', 'negative']) }),
    });
    return { reply: await ai.text(ctx, { prompt: `Reply to this ${sentiment} review: "${review}"` }) };
  },
});

Passing ctx ties the call to the step: it is cancelled when the step times out or the run is cancelled.

The model never picks the next step

The diagram decides the order. A model can write a field (score, isSpam) and a decision node can branch on it, but the path is always one drawn in the diagram.

Testing without a model

Replace the model call with your own function:

await run(flow, tools, input, {
  runAgent: async (request) => ({ bullets: ['a', 'b', 'c'] }),
});

Other ways to call a model

A tool is ordinary code. You can call any SDK or HTTP API directly inside run; you only lose the schema check and the automatic cancelling that ai.ask gives.