Claude Desktop on 3P works with Amazon Bedrock, Google Cloud, and Microsoft Foundry. The biggest difference is who processes conversation data. AWS and Google Cloud do this on their own platforms. On Microsoft Foundry, Anthropic does.
This guide is for admins of organizations on Claude Enterprise plans.
This article is part of Move from Claude Enterprise to Claude Desktop on a cloud provider, which defines the terms wave, overlap period, and cutoff date.
Topic | Amazon Bedrock | Google Cloud's Agent Platform (formerly Vertex AI) | Microsoft Foundry |
Setup guide | |||
Who processes conversation data | Google Cloud. Data handling is governed by Google Cloud. | Anthropic operates the models and handles conversation data as an independent processor for Microsoft, under both hosting options. | |
What reaches Anthropic from inference | The app doesn't send user data, prompts, or completions to Anthropic. AWS says model providers don't have access to prompts and completions. | The app doesn't send user data, prompts, or completions to Anthropic. For the models named in the data retention section of Google's abuse monitoring page, you must enable sharing prompts and responses with Anthropic. The same applies to models you enroll in the Cyber Verification Program. | Depends on the hosting option you choose. See data handling for Microsoft Foundry. |
Per-user sign-in | In-app AWS sign-in through IAM Identity Center | In-app Google sign-in, or Workforce Identity for Entra ID, Okta, Ping, and other IdPs | In-app Microsoft Entra ID sign-in by device code, browser, or OS identity broker |
Native web search | No. Add built-in search or a remote MCP search server. | Yes. Google turns it off by default through an organization policy. | Yes, on both hosting options, with nothing to configure in Claude Desktop. |
Where inference runs | Requests go to the regional endpoint you configure. The inference profile you list decides whether AWS processes them within one geography or in any commercial region. | A single region, the | Global Standard, or US Data Zone Standard for deployments hosted on Azure |
Public sector | In GovCloud regions, the app uses FIPS endpoints automatically. Confirm with AWS which models are in scope. For background, see our 2025 FedRAMP High and IL4/5 announcement. | Google says Claude through Agent Platform meets the FedRAMP High requirements. Confirm with Google which models are in scope. For background, see our 2025 FedRAMP High and IL2 announcement. | The public sector section of the overview applies to Amazon Bedrock and Google Cloud only. Confirm Azure Government availability with Microsoft. |
Running two identity providers (IdPs) at once | No. IAM Identity Center takes one identity source per AWS organization | Yes. A workforce pool can federate one or more IdPs | Per-user sign-in is Entra ID only. |
Limited. See the program article for who's eligible. | Available in preview. Accept the Advanced AI Safety Addendum, then enable advanced AI capabilities and data sharing for each model and location you use. | Available. Deploy a model, then link the subscription. After approval, turn on your tier in the Verification Portal. |
Note: Covered Models have their own retention rules. Their prompts and outputs are retained for at least 30 days by default on every platform where they're offered, and how you enable that retention differs by provider. Decide whether you need these models before you promise users a zero-retention setup.
For setup steps and provider-specific limits, use the setup guides in the first row of the table. Claude Desktop on 3P also supports an LLM gateway, Amazon Bedrock Mantle, and the Claude API. Most organizations won't need them for this move.
