Move from Claude Enterprise to Claude Desktop on a cloud provider

An Owner or Primary Owner of your Claude Enterprise organization. Only the Primary Owner can run an organization export.

An admin for the cloud account, project, or subscription that will run inference.

A way to deliver settings to devices: the Enterprise Admin Console, an MDM profile, or a bootstrap server.

The latest Claude Desktop release on each device. A device that moves from an earlier third-party configuration to the console should be on 5.11492.0 or later before its user signs in, because earlier releases migrate fewer items. Session export needs 4.23118.0 or later on both devices.

Each user moves their own history. Each user runs a one-time import on their own computer. Chats and projects come over, but uploaded files don't. See Import history from claude.ai.

Work is stored on the device. Conversation history, projects, and memory are stored on the user's computer. Prompts, responses, files, and tool outputs still go to your inference provider. There's no mobile app and no claude.ai web app. Claude runs only in the desktop app, on macOS and Windows (Linux is in beta). Users can't share chats or projects with each other. See the feature comparison.

Your cloud provider bills you. You pay for the tokens you use, and there are no seat licenses. See the feature comparison.

Your provider determines who processes your data. On Amazon Bedrock and Google Cloud, the cloud provider processes conversation data in the region or routing option you choose, and the app doesn't send user data, prompts, or completions to Anthropic. On Microsoft Foundry, Anthropic operates the Claude models as an independent processor for Microsoft. Some models add their own data retention and data sharing requirements. See Compare cloud providers.

Shared instructions and reference material: Package them as a skill or plugin and distribute it to everyone. See Distribute skills and plugins.

Shared files: Keep them in the file sharing or repository tools you already use. Claude writes artifacts and scheduled-task results to ~/Claude/ in the user's home directory. In Cowork sessions, files that Claude creates in a project are saved in the project's folders.

Global instructions: Go to Settings > Cowork and click "Edit" next to Global instructions. They're included in every session.

Project instructions: Add the text to a project's instructions. They apply to every session in that project.

Download any Word, Excel, PowerPoint, and PDF files Claude created for you, and export any artifacts you need in another format.

Gather the original files you uploaded to projects and chats. They don't come over.

Copy your memory and your Instructions for Claude (called Instructions in the newer Settings layout) into a local file.

Download the skills you built from Customize > Skills.

Make a note of the connectors you use, so you can ask for any that are missing.

Open each imported project. Review and accept its instructions, then add your files back. To use a project's folder in Chat, add the folder to the project again.

Open an imported chat and send a message. Click "Trust and resume" when asked.

Paste your saved memory and instructions into Claude Desktop.

Reconnect the connectors you use.

Run the import once more on your organization's cutoff date to pick up newer chats. It won't create duplicates.

Which parts of the app each group gets: Unified Claude, which combines Chat and Cowork, and Code.

Which permission modes users get. Agree on this with your security team before rollout. If users must approve every action, they may start to approve without reading, so decide which actions need a prompt. Decide on Auto mode, which new Code sessions normally start in unless you turn it off, bypass permissions mode, and which tools always ask.

How web search works. See Review web search and web fetch.

How to control cost: the in-app token cap, plus your provider's budgets, alerts, and quotas.

What telemetry to send to Anthropic and to your own collector.

How to handle data on devices: retention and legal hold, and session export for device changes. For a legal hold, set the chat, Cowork, and Code retention periods to 3650 days for the users on hold. The sessionRetentionHold setting stops working on December 17, 2026.

How to configure Claude Code CLI users.

When to make any IdP change: before the first wave moves, between waves, or after the last import.

What to do about restricted models and access programs: turn them on for the cloud account, keep some Claude Enterprise seats for that work, or stop using them.

Make the choices listed in Decisions to make before your pilot.

Name one owner each for the cloud account, device management, identity, network, and communications and training.

List what's in your Claude Enterprise organization: users, shared projects, skills your organization provides, and connectors in use.

List the users who rely on the Claude mobile apps or on claude.ai in a browser, such as executives who work from a phone or tablet. Claude Desktop on 3P has neither, so decide whether those users keep Claude Enterprise seats.

Agree on how you'll measure success, the order of your waves, and your cutoff date for claude.ai.

Count who uses the Claude Code CLI. Note the server-managed settings, cloud sessions, and claude.ai connectors they depend on.

List every sign-in your users will see, and note any IdP change planned for the same period. See Plan identity and sign-in.

List the access programs, restricted models, and custom roles you use today. See Plan for restricted models and access programs.

Enable or deploy the models with your provider, grant the inference role, and request enough quota for your users. Follow your provider's setup guide, linked in Compare cloud providers.

Connect your IdP to the cloud provider and test sign-in with a small group.

Run the Cowork readiness check on each hardware model. Cowork needs hardware virtualization and, on Windows, the Virtual Machine Platform feature.

Package the app and test automated deployment with the team that manages your devices. Allow time for this work. On Windows, use the .msix package, which Cowork requires. Installing it needs administrator rights.

Build the configuration in the console, or build it in the app and export it for MDM. Plan for the latest Claude Desktop release on every device.

Decide which settings users shouldn't change, and set each one before the first user signs in. For example, Code offers Auto mode unless you set autoModeEnabled to false. Deliver these settings from the console, an MDM profile, or a bootstrap server, because users can edit the local configuration file. The console alone doesn't keep a device on your settings. To enforce them, also deploy forceLoginOrgUUID in a machine-level managed profile. The key limits which organizations users can sign in to in Claude Desktop, so list your Enterprise Admin Console organization and every other organization they still need during the overlap period. Then check on a test device which settings users can still change. With MDM, confirm that the in-app configuration window opens read-only. On Windows, a misspelled value also locks the window, so compare each value in the diagnostic report's managed-config.txt with the value you set.

Open the required network hosts, and allow the app in your endpoint security tools by its signing identity.

Configure connectors, the plugin marketplace, web search, the banner, the token cap, and OpenTelemetry export.

Set up model access rules at the provider and in the app's model list.

Set budget alerts with your provider, and turn on cost attribution before the first user signs in.

For CLI users, deploy Claude Code managed settings with the provider variables, pinned models, model limits, and allowedProviders. See Move Claude Code CLI users.

If you hold an access program grant, link the cloud account in the Verification Portal and turn on access. Allow for the lead time.

Turn on member data export and the import wizard. See Turn on import.

Choose any optional import settings, such as the import prompt, the organization allowlist, and rollout by group.

Confirm that devices can reach the hosts that sign-in import needs.

Have the Primary Owner run an organization export for your records. The download link expires after 24 hours and may work only once, so download the file right away.

Send each wave the Checklist for users, along with your cutoff date.

Send developers the Checklist for developers, so they pull down cloud sessions before the cutoff date.

Postpone any changes to the claude.ai SSO connection and SCIM directory until the last import, or schedule the switch between waves.

With MDM or a bootstrap server, deploy the configuration first, then the app, so users open the app directly into Claude Desktop on 3P. With the console, save the configuration first, then have users install the app and sign in.

Check a device with Help > Troubleshooting > Generate Diagnostic Report. In the exported file, deployment-mode.txt shows whether the app is in third-party mode.

Pilot with a mix of roles. Where they apply, include a CLI developer, a user of each restricted model, and a user from each identity group.

Test the import, each connector, web search on each network users work from, the Primary Owner's history, and what users see when a request is flagged. Also test resuming a CLI session, cost reporting by user, and how spend appears in your provider's billing console.

Fix what the pilot finds and update your user guide. Then move each wave, with launch training and office hours.

For each wave, track how many users have installed the app, signed in, imported, and are still active in week two. The desktop_claude_ai_import event in your telemetry is logged for each claude.ai history import. Users can import more than once, so count distinct users. It's an info-level event, so set otlpDesktopLogLevel to info to receive it.

On the cutoff date, have users finish their work in claude.ai and run the import one last time. Do this before you block claude.ai on your network, because sign-in import needs it. If you use the console, don't block claude.ai, api.anthropic.com, or the hcaptcha.com hosts. Limit claude.ai and api.anthropic.com to your organizations with Tenant Restrictions instead. With MDM or a bootstrap server, the app also reads its update feed from claude.ai and api.anthropic.com by default. Before you block them, set updateViaUpdatesHost to true and allow releases.claude.com, or turn off auto-update and distribute new versions yourself. In Code sessions, Web Fetch checks each domain with api.anthropic.com and refuses the fetch if it can't, so also set skipWebFetchPreflight to true before you block that host. Keep admin access after the cutoff date, so you can run final exports and remove users.

Keep users in the new deployment. With MDM or a bootstrap server, set deploymentMode to 3p in the managed profile that sets your provider or bootstrap keys (Claude Desktop 4.16336.0 or later). Until every device has updated, also set disableDeploymentModeChooser to true in the profile and, with a bootstrap server, in its response. Don't set either key on devices that use the console. With the console, turn on Require this organization in Claude Desktop so members of both organizations must switch. That setting doesn't enforce sign-in. To enforce it, remove your Claude Enterprise organization's ID from forceLoginOrgUUID after the last wave has moved, except for users who keep Claude Enterprise.

For CLI users, narrow allowedProviders to the new provider, except on computers where users stay signed in to claude.ai in Claude Desktop, and retire your server-managed settings.

Control personal accounts on your domain and network with Restrict organization creation and Tenant Restrictions. If you use the console, include the new organization's ID in the allowed list.

Run a final organization export and audit log export. A removed user's data is still included in the Primary Owner's exports.

Some work may have to stay on Claude Enterprise, for example under an access program your cloud provider doesn't support yet, or for users who need the mobile apps. If so, keep those seats and reduce the size of the organization instead of closing it.

Remove users. With SCIM, users you remove from your IdP app are removed automatically. With JIT, they stay in your member list until an admin removes them. A user you remove only in Claude is added again the next time they log in with SSO. If your Enterprise Admin Console organization shares the SCIM directory, check how removing users from the IdP app affects it first. Removed users lose access immediately, but removal doesn't delete their account or data.

To clear stored data sooner, you can shorten your custom data retention period. Anything outside the new period is scheduled for permanent deletion as soon as you save, so finish your exports first. The minimum period is 30 days, and it doesn't cover the memory of users you've already removed. Chats in projects follow the project retention period instead, and artifacts published with a public link aren't deleted.

Before you close anything, confirm with your Anthropic account team which domain and SSO settings the new deployment still relies on. Then talk to them about ending the plan. To delete the organization, the Primary Owner sends a request to Support.

Claude Desktop on 3P overview

Feature comparison

Import history from claude.ai

Installation and setup

Deploy with Enterprise Admin Console

Configuration reference

Enterprise deployment overview for Claude Code

Export your organization's data

Switching to a different Identity Provider (IdP)

Cyber Verification Program

Covered Models

Migrate Claude Desktop from a cloud provider to Claude Enterprise