Providers

Configure OpenAI, OpenRouter, Anthropic, Gemini, Ollama, LM Studio, llama.cpp, Bedrock, or Vertex.

The SDK ships 9 built-in providers. Pick one (or several), set the env var, and pass a model id with the right prefix.

Provider summary

Provider	Env var(s)	Auth	Model id prefix
OpenAI	`OPENAI_API_KEY`	API key	`openai/gpt-4o-mini`, ...
OpenRouter	`OPENROUTER_API_KEY`	API key	`openai/gpt-4o-mini`, `anthropic/claude-...`, ...
Anthropic	`ANTHROPIC_API_KEY`	API key	`claude-sonnet-4-6`, `claude-opus-4-7`, ...
Gemini	`GEMINI_API_KEY`	API key	`google/gemini-2.0-flash-001`, ...
Ollama	none (local)	none	`ollama/llama3.2:3b`, ...
LM Studio	none (local)	none	`lmstudio/<model>`, ...
llama.cpp	none (local)	none	`llamacpp/<model>`, ...
AWS Bedrock	AWS credential chain or `AWS_BEARER_TOKEN_BEDROCK`	aws_bearer	`bedrock/us.anthropic.claude-sonnet-4-5-...`
GCP Vertex	`GOOGLE_APPLICATION_CREDENTIALS` or ADC	gcp_oauth	`vertex/anthropic/...` or `vertex/google/gemini-...`

OpenAI

export OPENAI_API_KEY="sk-..."

const agent = await Agent.create({
  apiKey: process.env.OPENAI_API_KEY,
  model: { id: "openai/gpt-4o-mini" },
});

OpenRouter

OpenRouter aggregates 300+ models behind one API. Cheapest path for prototyping.

export OPENROUTER_API_KEY="sk-or-v1-..."

const agent = await Agent.create({
  apiKey: process.env.OPENROUTER_API_KEY,
  model: { id: "openai/gpt-4o-mini" }, // any OpenRouter-supported model id
});

Anthropic

export ANTHROPIC_API_KEY="sk-ant-..."

const agent = await Agent.create({
  apiKey: process.env.ANTHROPIC_API_KEY,
  model: { id: "claude-sonnet-4-6" },
});

Gemini

export GEMINI_API_KEY="..."

const agent = await Agent.create({
  apiKey: process.env.GEMINI_API_KEY,
  model: { id: "google/gemini-2.0-flash-001" },
});

Ollama (local)

# Install + start Ollama: https://ollama.com
ollama serve &
ollama pull llama3.2:3b

const agent = await Agent.create({
  apiKey: "ollama", // any non-empty placeholder — Ollama ignores Authorization
  model: { id: "ollama/llama3.2:3b" },
});

Set OLLAMA_HOST=http://192.168.1.50:11434 to point at a remote box.

LM Studio (local)

LM Studio defaults to port 1234.

const agent = await Agent.create({
  apiKey: "lmstudio",
  model: { id: "lmstudio/<your-loaded-model>" },
});

Override with LMSTUDIO_HOST=http://localhost:1234.

llama.cpp (local)

llama.cpp server defaults to port 8080.

const agent = await Agent.create({
  apiKey: "llamacpp",
  model: { id: "llamacpp/<model>" },
});

Override with LLAMACPP_HOST=http://localhost:8080.

AWS Bedrock

Two auth paths. Either set a Bearer token directly:

export AWS_BEARER_TOKEN_BEDROCK="bedrock-api-key-..."

Or rely on the standard AWS credential chain (~/.aws/credentials, IAM role, env vars) plus the optional peer:

pnpm add @aws/bedrock-token-generator @aws-sdk/credential-providers

Then:

const agent = await Agent.create({
  apiKey: process.env.AWS_BEARER_TOKEN_BEDROCK ?? "__bedrock_lazy_token__",
  model: { id: "bedrock/us.anthropic.claude-sonnet-4-5-20250929-v1:0" },
});

Common gotcha: the AWS account needs to complete the "Anthropic use case details" form in the Bedrock Console (separate from "Model access"). Without it, every request returns ResourceNotFoundException.

GCP Vertex AI

Install the required peer:

pnpm add google-auth-library

Set up ADC:

gcloud auth application-default login
export GOOGLE_CLOUD_PROJECT="your-project-id"

Then pick a dialect:

// Claude on Vertex:
model: { id: "vertex/anthropic/claude-sonnet-4-5-20250929" }

// Gemini on Vertex:
model: { id: "vertex/google/gemini-2.0-flash-001" }

Multi-provider fallback

You can list a fallback chain — if the primary fails, the SDK tries the next:

const agent = await Agent.create({
  apiKey: process.env.OPENROUTER_API_KEY,
  model: { id: "openai/gpt-4o-mini" },
  providers: {
    routes: [{ provider: "openrouter" }],
    fallback: ["openai", "anthropic"],
  },
});

See the Agent concept for fallback semantics and the Credential pool for multi-key rotation.

Build your first real agent with tools, persistence, and memory.