Providers

A provider resolves model names into models. Every provider implements model.Provider, so agents and runners treat them the same way.

Provider	Package	Notes
OpenAI	`pkg/model/providers/openai`	Chat Completions API
Azure OpenAI	`pkg/model/providers/openai`	Same package, `SetAPIType(APITypeAzure)`
Gemini	`pkg/model/providers/openai`	OpenAI compatible endpoint, `NewGeminiProvider`
Anthropic	`pkg/model/providers/anthropic`	Messages API
Amazon Bedrock	`pkg/model/providers/bedrock`	Converse API, SigV4 signed
LM Studio	`pkg/model/providers/lmstudio`	Local models

OpenAI

provider := openai.NewProvider(os.Getenv("OPENAI_API_KEY"))
provider.SetDefaultModel("gpt-4o-mini")

provider.WithRateLimit(60, 150000)          // requests/min, tokens/min
provider.WithRetryConfig(3, 2*time.Second)  // retries with exponential backoff
provider.WithOrganization("org-...")

Azure OpenAI

provider := openai.NewProvider(os.Getenv("AZURE_OPENAI_API_KEY"))
provider.SetBaseURL("https://<your-resource>.openai.azure.com")
provider.SetAPIType(openai.APITypeAzure)     // or APITypeAzureAD for Entra ID tokens
provider.SetAPIVersion("2024-10-21")
provider.SetDefaultModel("<your-deployment-name>")

For Azure the model name is the deployment name. With APITypeAzure the key is sent as the api-key header; with APITypeAzureAD it is sent as a bearer token. See examples/azure_openai_example.

Gemini

Gemini exposes an OpenAI compatible endpoint, but it rejects several JSON Schema keywords that OpenAI accepts. The provider adapts tool schemas before sending them, which is enabled automatically for Gemini base URLs:

provider := openai.NewGeminiProvider(os.Getenv("GEMINI_API_KEY"))
provider.SetDefaultModel("gemini-2.5-flash")

The explicit form is:

provider := openai.NewProvider(os.Getenv("GEMINI_API_KEY"))
provider.SetBaseURL(openai.GeminiBaseURL)
provider.SetSchemaCompatibility(openai.SchemaCompatibilityStrict)

Schema compatibility modes

Mode	Behaviour
`SchemaCompatibilityDefault`	Safe normalization only: an object schema without properties is sent as “no parameters”, and `required` entries without a matching property are dropped
`SchemaCompatibilityStrict`	Additionally removes keywords strict endpoints reject: `$schema`, `$id`, `$ref`, `$defs`, `definitions`, `additionalProperties`, `patternProperties`, `const`, `default`, `examples`, `title`, `exclusiveMinimum`, `exclusiveMaximum`, `multipleOf`, `allOf`, `oneOf`, `not`, and unsupported `format` values
`SchemaCompatibilityNone`	Sends schemas untouched

The mode only affects what is sent over the wire; your tool definitions are never modified.

Anthropic

provider := anthropic.NewProvider(os.Getenv("ANTHROPIC_API_KEY"))
provider.SetDefaultModel("claude-sonnet-4-20250514")

provider.WithRateLimit(40, 80000)
provider.WithRetryConfig(3, 2*time.Second)

Amazon Bedrock

The Bedrock provider talks to the Converse API, which gives every model hosted on Bedrock - Anthropic Claude, Amazon Nova, Meta Llama, Mistral and others - the same request format including tool use.

provider := bedrock.NewProvider("eu-north-1")
provider.SetDefaultModel("anthropic.claude-3-5-sonnet-20241022-v2:0")

Credentials are read from AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY and AWS_SESSION_TOKEN, and the region falls back to AWS_REGION / AWS_DEFAULT_REGION. They can also be set explicitly:

provider.WithCredentials(accessKeyID, secretAccessKey, sessionToken)
provider.WithRegion("us-east-1")
provider.WithRetryConfig(5, time.Second)
provider.SetBaseURL("https://bedrock-runtime.eu-north-1.amazonaws.com") // VPC endpoints

Requests are signed with AWS Signature Version 4 using only the standard library, so the SDK stays dependency free. The model id may also be an inference profile ARN.

Streaming with Bedrock replays the completed answer as stream events, because the Bedrock streaming API uses a binary event stream format. The event sequence your code sees is the same as with the other providers.

See examples/bedrock_example.

LM Studio

provider := lmstudio.NewProvider()
provider.SetBaseURL("http://127.0.0.1:1234/v1")
provider.SetDefaultModel("gemma-3-4b-it")

Per-agent providers

An agent can carry its own provider, which is how a single workflow mixes providers. See multi-provider workflows.