Claude Platform on AWS: Enterprise Guide to Agents, Billing, Governance, and Bedrock Tradeoffs

Claude Platform on AWS: What enterprises need to know

Updated May 2026

TL;DR: Anthropic’s Claude Platform is now available through AWS Marketplace, letting teams use the native Claude experience—APIs, agent tooling, code execution, files, and more—directly from their AWS accounts with IAM-based access and Marketplace billing. It’s a fast way to get Anthropic Claude features under your cloud governance, but note Anthropic operates the service and processes requests outside the AWS security boundary, which matters for strict data-residency or in-cloud compliance needs.

What Claude Platform on AWS actually does

Claude Platform on AWS plugs Anthropic’s native Claude experience into your AWS account so your teams can sign in with AWS credentials, track spend in Cost Explorer, and audit calls through CloudTrail—without separate Anthropic accounts or invoices. It exposes the same APIs and tooling Anthropic ships directly while using AWS-native authentication, billing, and observability.

Quick definitions:

  • AWS Signature Version 4 (SigV4) — AWS’s standard request-signing method for authenticating API calls.
  • MCP (Model Connector Protocol) connector — Anthropic’s bridge for connecting Claude agents to external data sources and tools.

“AWS customers can use Anthropic’s Claude Platform through their AWS account without managing separate credentials, contracts, or invoices.”

Practical use cases that ship faster with this setup

  • Sales automation: Use Claude Managed Agents to orchestrate prospect research, email drafting, and CRM updates while billing AI consumption to centralized AWS accounts.
  • Internal developer copilots: Run Claude Code and code-execution workflows from developer workspaces for faster onboarding and fewer security exceptions.
  • Customer support & document workflows: Ingest files via the files API and run agent-driven triage or summarization pipelines tied to workspace-level access control.

Core features available through AWS

Major Claude Platform capabilities are available when accessed via AWS: Messages API, Claude Managed Agents (beta), advisor tool (beta), web search and web fetch, the MCP connector (beta), Agent Skills (beta), code execution, and files API (beta). These let teams build AI agents and automation pipelines using Anthropic-native primitives while managing access and spend in AWS.

How the AWS integration works — authentication, billing, and audit

Authentication

Two options: AWS IAM (SigV4) and API keys. For production, use temporary IAM credentials (STS) to enforce least privilege and simplify rotation. Workspaces are represented as IAM resources—your IAM policies reference a workspace ARN to grant access to a specific project or team.

IAM pattern (conceptual): assign principals (users/roles) permission to the workspace ARN, e.g. arn:aws:claude:<region>:<account-id>:workspace/<workspace-id>, and scope actions to only what’s needed (Messages:Send, Agents:Invoke, Files:Read).

Billing

Usage is billed through AWS Marketplace on a consumption basis, so all Claude spending appears with other AWS charges. Use resource tags and Cost Explorer to allocate and track costs per team, environment, or project.

Audit & observability

CloudTrail captures management events by default. You can enable data-event logging to record inference-level activity. Combine CloudTrail with your existing telemetry (CloudWatch, X-Ray, third-party SIEM) to track latency, failures, and request volumes.

Important operational nuance: processing outside the AWS security boundary

Anthropic operates the service and processes requests outside the AWS security boundary. Practically, that means:

  • Network hops: inference requests leave your AWS VPC and travel to Anthropic-managed endpoints, which can add latency compared with in-cloud model hosting.
  • Audit & compliance scope: while CloudTrail captures control-plane events, the underlying inference environment and data handling are governed by Anthropic’s operational controls—not AWS’s in-cloud certifications.
  • Data residency: if your policy requires that data and models remain physically inside your AWS tenant or specific geographic boundary, Amazon Bedrock (models hosted inside AWS) may be the better fit.

For regulated applications, ask Anthropic for their compliance docs, retention policies, and data-flow diagrams before committing to production.

Claude Platform on AWS vs Amazon Bedrock — a quick comparison

  • Hosting boundary: Claude on AWS = Anthropic-operated (outside AWS boundary); Bedrock = models hosted inside AWS.
  • Feature parity: Claude Platform provides Anthropic-native tooling and Managed Agents; Bedrock exposes foundation models via AWS APIs and integrates with AWS services natively.
  • Governance & billing: Both can appear in your AWS billing; Claude uses Marketplace for consumption billing while Bedrock is an AWS-native model hosting service.
  • When to pick which: Choose Claude Platform on AWS when you need Anthropic’s agent tooling and expedited vendor innovation; choose Bedrock when strict in-cloud residency, certifications, or network isolation are mandatory.

A decision matrix for common buyer profiles

  • Startup POC — Goal: speed. Recommendation: Claude Platform on AWS for rapid feature access and simplified billing.
  • Enterprise pilot — Goal: governance + vendor features. Recommendation: Claude Platform on AWS with workspace isolation and CloudTrail enabled; run compliance checks early.
  • Regulated finance/healthcare — Goal: strict residency & certified controls. Recommendation: Evaluate Bedrock first; if Anthropic features are required, request SOC/ISO evidence and contractual protections before using Claude Platform on AWS.

Quickstart checklist for pilots

  1. Activate Claude Platform on AWS from AWS Marketplace and link it to the appropriate payer account.
  2. Create a workspace per team or environment and capture its workspace ARN for IAM policies.
  3. Use temporary IAM credentials (STS) for test roles to avoid long-lived secrets.
  4. Run a controlled pilot: 1–2 agent workflows, a file ingestion pipeline, and a code-execution test to validate tooling.
  5. Enable CloudTrail data events (if you need inference logs) and tag all resources for cost tracking.
  6. Gather compliance artifacts from Anthropic (data handling, retention, certifications) and run a gap analysis against your requirements.

Cost & operational considerations

There’s no one-size-fits-all price. To model costs, estimate:

  • Calls per minute / peak throughput
  • Average context length and output tokens per call (affects inference units)
  • Frequency and size of file storage or retrievals
  • Agent orchestration overhead (tooling calls, fetches, multistep workflows)
  • Marketplace fees or procurement overhead

During pilots capture telemetry: end-to-end latency, error rates, request and token counts, and CloudTrail event volumes (data-event logging can produce high log volume). Use those metrics to project 30/90/365-day costs in Cost Explorer.

Governance & compliance questions to ask Anthropic

  • Where are inference requests processed and which regions are used?
  • What certifications (SOC 2, ISO 27001, HIPAA, etc.) does the Anthropic environment hold?
  • What is the data retention policy for inference logs and uploaded files?
  • Can Anthropic provide a data-flow diagram and a list of subprocessors?
  • What SLAs and support channels apply to Marketplace customers using Claude Platform on AWS?

When not to use Claude Platform on AWS

It’s not the right fit if your policies require:

  • All processing to remain inside your AWS tenant or specific geographic boundary.
  • Only using models in an environment covered by AWS-internal certifications for in-scope systems (e.g., certain FISMA/HIPAA setups without vendor attestations).
  • Ultra-low-latency inference where even the extra network hop to a vendor-operated endpoint is unacceptable.

FAQ

  • Is Claude Platform hosted in AWS?

    No—the offering is accessible via your AWS account, but Anthropic operates the inference environment and processes requests outside the AWS security boundary.

  • How is billing handled?

    Usage is billed through AWS Marketplace on a consumption basis and appears on your AWS invoice.

  • Does it support IAM?

    Yes. AWS IAM (SigV4) is supported and recommended; workspaces are represented as IAM resources and can be scoped by ARN.

  • Can I get inference-level logs?

    Yes—enable CloudTrail data events to capture inference activity, and route logs to CloudWatch or your SIEM for analysis.

  • How do I choose between Claude on AWS and Amazon Bedrock?

    If Anthropic-native tooling and Managed Agents are critical and strict in-cloud residency isn’t required, choose Claude Platform on AWS; otherwise, prefer Bedrock for in-cloud model residency and isolation.

Suggested next steps for architects and product leaders

  • Run a focused pilot using one or two high-impact workflows (e.g., sales email automation or a developer copilot).
  • Collect telemetry (latency, token usage, CloudTrail events) and a 90-day cost estimate.
  • Request compliance artifacts and SLA details from Anthropic before scaling production.
  • Use workspace ARNs and temporary IAM credentials to enforce least privilege and simplify audits.

Want a ready-made comparison and pilot checklist? Download our one‑page Claude vs Bedrock decision matrix and pilot checklist at /resources/claude-vs-bedrock-matrix or read more on related topics at AI for business, AI agents, and AI for sales.

Briefed by Dani Mitchell, Sofian Hamiti, Eugenio Soltero, Antonio Rodriguez, and Ayan Ray from AWS generative AI and product teams.

0