Skip to content

providers.md

Latest commit

 

History

History
88 lines (64 loc) · 2.37 KB

providers.md

File metadata and controls

88 lines (64 loc) · 2.37 KB
Raw

Model Providers

An oya Agent takes a model. oya ships three provider adapters - Anthropic, OpenAI, and Google - each a small function that returns a LanguageModel. They're interchangeable: the rest of your agent (tools, instructions, generate / stream) stays identical no matter which one plans.

Each provider is a separate entry point, so you only pull in what you use:

import { anthropic } from "oyadotai/anthropic";
import { openai } from "oyadotai/openai";
import { google } from "oyadotai/google";

Anthropic

Reads ANTHROPIC_API_KEY from the environment unless you pass apiKey.

import { Agent, createTool } from "oyadotai";
import { anthropic } from "oyadotai/anthropic";

const agent = new Agent({
  model: anthropic("claude-haiku-4-5-20251001"),
  tools: { get_weather: getWeather },
});

const { text } = await agent.generate("How's the weather in NYC?");

OpenAI

Reads OPENAI_API_KEY from the environment unless you pass apiKey.

import { Agent } from "oyadotai";
import { openai } from "oyadotai/openai";

const agent = new Agent({
  model: openai("gpt-4o"),
  tools: { get_weather: getWeather },
});

Google

Reads GEMINI_API_KEY (or GOOGLE_API_KEY) from the environment unless you pass apiKey.

import { Agent } from "oyadotai";
import { google } from "oyadotai/google";

const agent = new Agent({
  model: google("gemini-2.5-pro"),
  tools: { get_weather: getWeather },
});

Passing the key explicitly

Every provider accepts an options object with apiKey, so you don't have to rely on environment variables - useful in serverless or multi-tenant setups:

anthropic("claude-haiku-4-5-20251001", { apiKey: process.env.MY_ANTHROPIC_KEY });
openai("gpt-4o", { apiKey: myKey });
google("gemini-2.5-pro", { apiKey: myKey });

If no key is found, the provider throws a clear error naming the environment variable to set.

Swapping providers

Because the provider is just the model value, switching is a one-line change - your tools, instructions, and the plan-once execution model are unchanged:

- model: anthropic("claude-haiku-4-5-20251001"),
+ model: openai("gpt-4o"),

The model only ever emits the plan; from there the runtime executes the DAG the same way regardless of which provider planned it. See Projection Types for what the model does and doesn't get to see.