Anthropic’s AWS SDK Clients Make Claude Platform on AWS Feel Like a First-Class API
The least glamorous part of an enterprise AI launch is usually the most important one: the SDK import. Press releases say a platform is generally available. Procurement hears “billing and contracts are easier now.” Engineers ask a different question: can I call this thing from Python and TypeScript without hand-rolling authentication middleware that becomes everyone’s favorite future outage?
Anthropic answered that question this week with AWS-specific SDK clients for Claude Platform on AWS. The Python SDK v0.101.0 adds AnthropicAWS and AsyncAnthropicAWS. The TypeScript AWS SDK v0.3.0 adds AnthropicAws. Both releases were published May 11 around 15:46 UTC, shortly after AWS and Anthropic made the bigger commercial pitch: Claude Platform is now generally available through AWS accounts.
The headline sounds redundant until you look closely. AWS already has Bedrock. Anthropic already has a first-party API. Claude Platform on AWS is a third path: Anthropic’s native platform experience, exposed through AWS commercial rails. AWS provides IAM access control and Marketplace billing integration. Anthropic operates the platform and inference path. That distinction is not packaging trivia. It decides which API your engineers use, which features arrive first, which compliance envelope applies, which identity system governs access, and which incident page people stare at when a session gets weird.
This is not Bedrock with a different logo
Bedrock is the AWS-native model service. It is the conservative enterprise answer: AWS APIs, AWS service boundary, Bedrock governance, and the operational expectations that come with that. Claude Platform on AWS is different. AWS’s own documentation says Claude Platform on AWS uses Anthropic’s Messages API directly and aims for feature parity with the first-party Claude API except for listed gaps. That is the point. This is the “AWS procurement, Anthropic platform” compromise.
The tradeoff is explicit. AWS says customer data is processed outside the AWS security boundary. The docs say data may not reside in AWS, inference may route to Anthropic’s primary cloud, and PrivateLink protects only the path from a VPC to the AWS endpoint before requests continue to Anthropic infrastructure. Teams with HIPAA requirements are told to evaluate Claude in Amazon Bedrock instead. Teams with strict residency requirements need to inspect request-level inference geography behavior, not assume the AWS logo turns every boundary green.
In exchange, the feature surface is much closer to Anthropic’s fast-moving developer platform. AWS lists support for Claude Managed Agents beta, advisor strategy beta, web search, web fetch, code execution, Files API beta, Skills beta, MCP connector beta, prompt caching, citations, batch processing, and the Claude Console. That is the value proposition: keep AWS billing and IAM, but avoid waiting for every new Anthropic platform feature to be repackaged into a slower cloud-native service surface.
The gaps matter too. AWS docs list unavailable or limited features including HIPAA readiness, Priority Tier, several Admin API endpoints, spend limits, Claude Code workspace and Analytics API, OAuth, OpenAI-compatible endpoints, and workspace-level inference geography controls. That is not a footnote. For a regulated team, those absences may decide the architecture before anyone benchmarks a model.
SigV4 is where adoption dreams go to lose a week
The SDK releases matter because they remove the ugly adoption tax: AWS request signing. SigV4 is perfectly reasonable infrastructure, but no product team should be implementing its own body hashing, header canonicalization, credential-chain resolution, workspace ID routing, region-derived base URLs, and proxy edge cases just to send a prompt.
The Python client leans on Botocore SigV4Auth with service name aws-external-anthropic and a base URL template of https://aws-external-anthropic.{region}.api.aws. It resolves auth mode in a clear order: constructor api_key means API key mode; explicit AWS access and secret keys mean SigV4; aws_profile means SigV4; ANTHROPIC_AWS_API_KEY means API key; and then the default AWS credential chain gets a turn. Region comes from the constructor or AWS_REGION/AWS_DEFAULT_REGION. Workspace comes from the constructor or ANTHROPIC_AWS_WORKSPACE_ID. It even removes the connection header from the signed header set because proxies may strip it. That is the kind of tiny compatibility fix engineers only appreciate after losing an afternoon to it.
The TypeScript SDK mirrors the same shape with apiKey, awsAccessKey, awsSecretAccessKey, awsSessionToken, awsProfile, custom providerChainResolver, workspaceId, and skipAuth. It uses Smithy and AWS SDK primitives including @smithy/signature-v4, @aws-sdk/credential-providers, @smithy/config-resolver, and @aws-crypto/sha256-js. It also adds a ready promise because region and base URL resolution may load asynchronously from ~/.aws/config. That is a good API design choice: fail fast on configuration instead of discovering the problem on the first production request.
The presence of skipAuth is pragmatic and mildly terrifying, as these things usually are. It is useful when an internal gateway handles authentication. It also means the security boundary has moved out of the SDK and into infrastructure configuration. Teams should allow it only where a reviewed proxy enforces identity, workspace scoping, logging, and request policy. “The gateway handles it” is not a control. It is a sentence that requires evidence.
IAM becomes your Claude governance layer
The workspace model is where engineering leaders should slow down. On first-party Claude, workspace membership is a Claude-side concept. On Claude Platform on AWS, AWS docs describe IAM policies granting aws-external-anthropic actions against workspace ARNs. That fits organizations already living in IAM, SCPs, permission boundaries, resource tags, and access reviews. It also means Claude access inherits whatever mess exists in your AWS identity posture.
If your AWS roles are clean, tagged, least-privilege, and reviewed, this is a win. If your AWS roles are a historical dig site with production-admin fossils in every layer, now your Claude governance is standing on the same archaeology. The good news is that CloudTrail, IAM policy review, Marketplace billing, and cost allocation tags are familiar enterprise muscles. The bad news is that people often overestimate how strong those muscles are.
Claude Managed Agents on AWS add another operational wrinkle. AWS docs describe agents, sessions, environments, credential vaults, and memory stores as IAM resources, and autonomous sessions require re-authentication every six hours. That is exactly the kind of limit that sounds reasonable in documentation and becomes a 2 a.m. “why did the agent stop?” incident when nobody owns credential refresh. Long-running agent workflows are infrastructure. Treat them that way.
The right rollout is not “change the import and ship.” Start with a tiny Python and TypeScript smoke test using both API-key and SigV4 auth. Verify region and workspace resolution. Confirm CloudTrail events. Test IAM scoping against workspace ARNs. Decide whether skipAuth is allowed anywhere. Validate cost allocation before a team runs a batch job that turns into an accounting scavenger hunt. Document when to use Bedrock, first-party Claude, and Claude Platform on AWS: regulated data, fast agent features, batch jobs, managed-agent sessions, and Claude Code-related workflows will not all have the same answer.
The practitioner take is simple: Claude Platform on AWS is not “Claude on Bedrock, but again.” It is Anthropic’s fast-moving developer platform entering the AWS procurement channel. The SDK clients are the tell. Anthropic wants teams to use the native Messages API and agent/tooling surface from AWS accounts without every application becoming a SigV4 project. That is a good thing. It is also a governance tradeoff, not a compliance magic trick.
Sources: Anthropic Python SDK v0.101.0, Anthropic TypeScript AWS SDK v0.3.0, Claude Platform on AWS announcement, AWS GA announcement, AWS Claude Platform user guide, AWS feature support docs, Hacker News discussion