Skip to content

Anthropic & Bedrock

DeepAgentsAnthropic provides two ChatModel implementations:

  • AnthropicChatModel - calls api.anthropic.com directly (or any Anthropic-compatible endpoint) via the Messages API over SSE.
  • BedrockChatModel - calls AWS Bedrock's invoke-with-response-stream endpoint using SigV4 request signing or a bearer token (Amazon Bedrock API key). Reuses the same Anthropic message codec, so behaviour is identical from the agent's perspective.

AnthropicChatModel

public struct AnthropicChatModel: ChatModel {
    public init(
        baseURL: URL,
        model: String,
        apiKey: String? = nil,
        ...,
        parameters: AnthropicGenerateParameters = .init(),
        anthropicVersion: String = "2023-06-01",
        betaHeaders: [String] = [],
        transport: (any AnthropicStreamingTransport)? = nil
    )
}
Parameter Type Default Purpose
baseURL URL (required) Root URL; the adapter POSTs to {baseURL}/v1/messages
model String (required) Model identifier, e.g. "claude-opus-4-8"
apiKey String? nil Sent as x-api-key; read from environment when nil
parameters AnthropicGenerateParameters .init() Sampling parameters (temperature, max tokens, etc.)
anthropicVersion String "2023-06-01" Value of the anthropic-version header; matches the stable dated API version
betaHeaders [String] [] Values appended to anthropic-beta; use to opt into preview features
transport (any AnthropicStreamingTransport)? nil Inject a custom transport for testing or proxying

The adapter sends a standard POST {baseURL}/v1/messages with:

  • x-api-key: <key>
  • anthropic-version: <anthropicVersion>
  • anthropic-beta: <betaHeaders joined by comma> (omitted when empty)
  • content-type: application/json
  • accept: text/event-stream

Responses are consumed as SSE; the AnthropicDecoder extracts tool use blocks, reasoning content, and text content into AgentMessage fields.

Error handling

AnthropicModelError is thrown on non-2xx responses or malformed SSE frames.

BedrockChatModel

BedrockChatModel targets the Bedrock invoke-with-response-stream endpoint. It authenticates with either AWS SigV4 request signing (SigV4Signer, built on CryptoKit) or an Amazon Bedrock API key (bearer token), selected via BedrockAuth, and decodes the AWS binary event-stream framing around the same Anthropic SSE payload.

public struct BedrockChatModel: ChatModel {
    public init(
        region: String,
        model: String,
        auth: BedrockAuth,
        baseURL: String? = nil,   // endpoint base, used verbatim; required for bearer-token auth
        ...
    )
}

BedrockAuth

public enum BedrockAuth: Sendable, Equatable {
    case sigV4(BedrockCredentials)
    case bearerToken(String)

    public static func resolve(bearerToken: String? = nil) -> BedrockAuth?
}

resolve(bearerToken:) prefers a bearer token over SigV4: an explicit bearerToken, then the AWS_BEARER_TOKEN_BEDROCK environment variable, then BedrockCredentials.fromEnvironment(); it returns nil when none are set. Bearer auth sends Authorization: Bearer <token> and requires a baseURL (used verbatim); SigV4 derives the endpoint from region.

BedrockCredentials

public struct BedrockCredentials: Sendable {
    public let accessKey: String
    public let secretKey: String
    public let sessionToken: String?

    public static func fromEnvironment() -> BedrockCredentials?
}

fromEnvironment() reads AWS_ACCESS_KEY_ID, AWS_SECRET_ACCESS_KEY, and optionally AWS_SESSION_TOKEN from ProcessInfo.processInfo.environment. Returns nil if the required keys are absent.

Cross-region inference profiles

The model parameter accepts cross-region inference profile IDs of the form us.anthropic.claude-.... Use these when your Bedrock quota is spread across regions and you want Bedrock to route automatically. Pass the profile ID exactly as shown in the Bedrock console.

Error handling

BedrockModelError is thrown on signing failures, HTTP errors, or malformed event-stream frames.

Examples

import DeepAgents
import DeepAgentsAnthropic

let model = AnthropicChatModel(
    baseURL: URL(string: "https://api.anthropic.com")!,
    model: "claude-opus-4-8",
    apiKey: ProcessInfo.processInfo.environment["ANTHROPIC_API_KEY"]
)

let agent = createDeepAgent(
    model: model,
    systemPrompt: "You are a helpful assistant.",
    maxIterations: 24
)

let ok = await agent.run([.human("What is 2 + 2?")]) { event in
    // handle AgentEvent
}

import DeepAgents
import DeepAgentsAnthropic

// Prefers a bearer token (AWS_BEARER_TOKEN_BEDROCK) over SigV4 env credentials
// (AWS_ACCESS_KEY_ID / AWS_SECRET_ACCESS_KEY / AWS_SESSION_TOKEN).
guard let auth = BedrockAuth.resolve() else {
    fatalError("No Bedrock credentials: set AWS_BEARER_TOKEN_BEDROCK or AWS_ACCESS_KEY_ID/SECRET")
}

let model = BedrockChatModel(
    region: "us-east-1",
    // Cross-region inference profile - Bedrock routes across us-* regions automatically
    model: "us.anthropic.claude-opus-4-8-20251101-v1:0",
    auth: auth,
    baseURL: "https://bedrock-runtime.us-east-1.amazonaws.com" // required for bearer-token auth
)

let agent = createDeepAgent(
    model: model,
    systemPrompt: "You are a helpful assistant.",
    maxIterations: 24
)