llm-core

Appearance

Your First Workflow ​

This guide will show you how to build a production-ready LLM workflow in less than 20 lines of code.

We'll build a tool-calling agent that uses OpenAI, then we'll swap it to Anthropic without changing any business logic.

1. The Blank Slate ​

Start by importing the recipes handle. This is your entry point for authoring.

ts
import { recipes } from "@geekist/llm-core/recipes";
import type { AgentRecipeConfig } from "@geekist/llm-core/recipes";

// Example: keep a typed config around
const config: AgentRecipeConfig = {};

2. Pick a Recipe ​

Recipes are declarative flows. We use recipes.*() to start one. You can define your own steps later, or load a standard recipe like "agent".

ts
import { recipes } from "@geekist/llm-core/recipes";

// "agent" is a standard recipe for loop-based agents.
const agent = recipes.agent();

type AgentRecipeConfig = Parameters<typeof agent.configure>[0];

3. Add the Brain (Adapters) ​

Recipes are abstract. They need Adapter Plugs to connect to the real world. The "agent" flow expects a model adapter.

> **Adapters** are the bridge between your code and the AI ecosystem (OpenAI, Anthropic, LangChain, etc.).

ts
import { recipes } from "@geekist/llm-core/recipes";
import { fromAiSdkModel } from "@geekist/llm-core/adapters";
import type { Model } from "@geekist/llm-core/adapters";
import { openai } from "@ai-sdk/openai";

const model: Model = fromAiSdkModel(openai("gpt-4o"));

// Configure the recipe with an OpenAI adapter
const workflow = recipes.agent().defaults({ adapters: { model } }).build();

4. Run It ​

Build the workflow and run it with an input.

ts
// ... imports

const result = await workflow.run({
  input: "What is the capital of France?",
});

if (result.status === "ok") {
  const answer: string | undefined = result.artefact.answer;
  console.log(answer); // "Paris"
}

5. The "Aha" Moment: Swapping Providers ​

Your boss wants to switch to Anthropic? No problem. You don't need to rewrite your agent logic. Just swap the adapter.

diff
- import { openai } from "@ai-sdk/openai";
+ import { anthropic } from "@ai-sdk/anthropic";

- const model = fromAiSdkModel(openai("gpt-4o"));
+ const model = fromAiSdkModel(anthropic("claude-3-5-sonnet-20240620"));

const workflow = recipes.agent().defaults({ adapters: { model } }).build();

That's it. Same inputs, same outputs, different brain.

TIP

Curious about what other adapters exist? See the Adapter API Reference.

Next Steps ​

Now that you've built your first agent, let's understand how data actually flows through it.