Claude Cowork Enterprise: EU Setup on Bedrock or Google

TL;DR

Claude Cowork enterprise deployments can run in third-party (3P) mode: Claude Desktop sends every Cowork, Chat and Code request to your own Amazon Bedrock, Google Cloud or gateway endpoint and keeps history on the device. With the Bedrock EU profile or the Google eu region, inference stays in the EU. You pay tokens to the cloud provider, no seats.

What is Claude Cowork and what does 3P mode change?

Cowork is the agentic workspace inside Claude Desktop: it plans multi-step tasks, works in local files and produces spreadsheets, decks and documents. In 3P mode, the same app sends inference to a provider you configure instead of Anthropic, loads its web UI from the app bundle and stores conversations on the user’s disk.

According to the Cowork overview, Cowork uses the same agentic architecture as Claude Code, without a terminal. On the standard plans it is available on Pro, Max, Team and Enterprise. Anthropic calls the enterprise variant “Claude Desktop on third-party (3P)”, and the old cowork/3p doc URLs now redirect there.

The 3P overview lists five inference targets: Google Cloud’s Agent Platform (formerly Vertex AI), Amazon Bedrock, Microsoft Foundry, an LLM gateway you operate, or the Anthropic API. The architecture differences:

Component	Standard Claude Desktop	Claude Desktop on 3P
Model inference	Anthropic API	Your Bedrock, Google Cloud, Foundry or gateway endpoint
Web application	Loaded from claude.ai	Bundled inside the desktop app
User identity	Anthropic account	Local device identity (or Anthropic account if managed from the admin console)
Conversation storage	Anthropic backend	Local disk on the user’s machine
Code execution sandbox	Local VM	Local VM (identical)
Configuration	claude.ai admin console	MDM profile, bootstrap server or claude.ai admin console

Anthropic is explicit about the target group: organizations that cannot send data through claude.ai or the Claude API, including “international enterprises” that require in-region residency and cannot send conversation data to the United States. The same page also says that if you can use the first-party products, Cowork on a Team or Enterprise plan is simpler and gets new features faster. We agree: 3P is a compliance tool, not a default.

Which Cowork features are missing on 3P?

Most of Cowork works on 3P: Chat, Cowork, Code, projects, MCP connectors, skills, plugins, plugin marketplaces, artifacts, memory and scheduled tasks. What you lose is the cloud side of claude.ai: mobile, web access, voice mode, sharing, Claude Design, the Compliance API and the Analytics API.

This contradicts a claim circulating in video tutorials that 3P mode has “no Chat and no skill marketplace”. The current feature matrix shows both Chat and plugin marketplaces as available. What is actually missing:

Feature	Claude Enterprise	Claude Desktop on 3P
Chat, Cowork, Code	Yes	Yes
Skills, plugins, hooks, plugin marketplaces	Yes	Yes
Web search	Yes	Yes, but Bedrock and plain gateways need a search provider
Memory	Yes (Anthropic)	Yes, stored on the device
Project and plugin sharing	Yes	No
Mobile, claude.ai web, voice mode	Yes	No
Claude Design, Claude Security, Claude Tag	Yes	No
Compliance API, Analytics API	Yes	No (use OpenTelemetry)
Per-user usage caps, retention policies, OpenTelemetry	Yes	Yes

Two details matter for EU teams. First, web search on Bedrock is not native: the web tools page tells you to add the bundled websearch server with Brave, Tavily, Exa or your own backend, which is a new processor for your records. Second, audit evidence comes from your own OpenTelemetry collector (otlpEndpoint), which exports metadata by default and prompt content only as an explicit opt-in.

How do you keep Cowork data in the EU?

Point Cowork at an EU endpoint and keep devices in the EU. On Bedrock, list EU inference profiles such as eu.anthropic.claude-sonnet-5 and set an EU region. On Google Cloud, set the region to eu. Microsoft Foundry offers no EU data zone for Claude, so it does not meet a strict EU requirement today.

Anthropic defines residency on 3P by two factors: the cloud region you select, and the physical location of the device, because conversations are persisted there. For Bedrock and Google Cloud, requests go directly from the device to the regional endpoint; conversation data goes only there, to local disk, and optionally to your OpenTelemetry collector.

Provider	EU setting in Cowork	What stays in the EU	Caveat
Amazon Bedrock	inferenceBedrockRegion = e.g. eu-central-1, models eu.anthropic.claude-sonnet-5, eu.anthropic.claude-opus-5	Inference within EU regions (AWS: “Keeps data within EU regions”)	Fable 5.1 has no EU profile; Frankfurt is only the entry point
Bedrock Mantle	inferenceProvider = mantle	Single region: Ireland or Stockholm	Separate endpoint, see our Bedrock guide
Google Cloud	inferenceVertexRegion = eu	EU member states only (UK and Switzerland excluded)	Private Google Access not supported for multi-region; use Private Service Connect
Microsoft Foundry	Deployment type in Foundry	Nothing guaranteed	Global Standard or US Data Zone only; Anthropic operates the models

Details on the Bedrock profiles and SCP pitfalls are in our Claude on AWS Bedrock EU guide, the Google side in Claude on Vertex AI in Europe. Both apply unchanged, because Cowork on 3P simply calls the same endpoints. For the platform itself, see our Amazon Bedrock guide.

Anthropic is not fully out of the picture. The app still contacts Anthropic hosts for crash reports, product analytics, favicons and artifact previews, auto-updates and the model catalog. Each can be switched off. After that, telemetry and egress says only downloads.claude.ai remains, for the VM bundle and CLI binary at session start. Anthropic states that telemetry never contains prompt or response content. For the processor question, the app licence falls under Anthropic’s Commercial Terms, while inference is “billed by, and subject to your agreement with” the cloud provider (legal page).

How do you set up Cowork on Bedrock step by step?

Enable the Claude models in your EU region, grant users Bedrock invoke rights via IAM Identity Center, test the configuration on one machine through Developer mode, then export it as a macOS or Windows profile for your MDM. Budget half a day for a pilot if IAM Identity Center already exists.

1. Enable models. In the Bedrock console, request access to the Claude models in the same region you will set as inferenceBedrockRegion. Access is granted per region.
2. Create a permission set in IAM Identity Center that allows bedrock:InvokeModel and bedrock:InvokeModelWithResponseStream. Anthropic recommends a session duration of 8 to 12 hours.
3. Pick authentication. The Bedrock deployment page offers a bearer token (proof of concept only, shared key), in-app AWS sign-in via Identity Center (inferenceBedrockSso*, per-user CloudTrail identity, no AWS CLI) or a named AWS profile.
4. Test on one device. Install Claude Desktop, do not sign in, then go to Help > Troubleshooting > Enable Developer Mode and Developer > Configure Third-Party Inference. Enter the values, click Apply Changes and Save & Restart.
5. Harden and export. Add retention, telemetry and egress keys, click Export and deploy the .mobileconfig or .reg file via Jamf, Intune or Group Policy.
6. Verify. Help > Troubleshooting > Generate Diagnostic Report shows in managed-config.txt which keys were applied. Misspelled keys are silently ignored.

A minimal EU configuration, written as plain JSON (the form a bootstrap server returns; the in-app Export converts it for macOS and Windows). All key names are from the configuration reference:

{
  "inferenceProvider": "bedrock",
  "inferenceBedrockRegion": "eu-central-1",
  "inferenceBedrockSsoStartUrl": "https://d-xxxxxxxxxx.awsapps.com/start",
  "inferenceBedrockSsoRegion": "eu-central-1",
  "inferenceBedrockSsoAccountId": "123456789012",
  "inferenceBedrockSsoRoleName": "ClaudeCoworkEU",
  "inferenceModels": ["eu.anthropic.claude-sonnet-5", "eu.anthropic.claude-opus-5"],
  "coworkSessionRetentionDays": 90,
  "chatSessionRetentionDays": 90,
  "disableNonessentialTelemetry": true,
  "disableNonessentialServices": true,
  "otlpEndpoint": "https://otel-collector.example.com:4318"
}

Before rollout, confirm where the EU profile actually routes from your source region:

aws bedrock get-inference-profile --region eu-central-1 --inference-profile-identifier eu.anthropic.claude-sonnet-5 --query "models[].modelArn" --output table

On Google Cloud, replace the provider block with inferenceProvider = vertex, inferenceVertexProjectId and inferenceVertexRegion = eu, and grant users the Vertex AI User role. If developers also use the Claude Code CLI, the same EU logic is covered in Claude Code on Bedrock with EU residency.

Can Cowork run on local models with Ollama or LM Studio?

Not with Anthropic’s support. Anthropic’s 3P docs describe gateways for “Claude models” only and never mention Ollama or LM Studio. Ollama itself documents a Claude Desktop integration on macOS that lists Cowork as supported. Treat it as an unsupported experiment, not as an enterprise deployment.

The technical route exists because the gateway provider only needs an Anthropic-compatible POST /v1/messages endpoint with streaming and tool use (gateway page). Ollama and LM Studio both expose such an endpoint. The Ollama Claude Desktop page sets this up from the Ollama app, notes that Windows support is “coming soon”, and says cloud models require an Ollama sign-in and sometimes a paid plan. LM Studio documents Claude Code, but we found no Claude Desktop or Cowork page.

Three reasons we advise against it for production:

• Model discovery is built for Claude. Anthropic states that auto-discovery shows only models whose IDs are recognizably Claude. The tutorials work around this by renaming open-weight models to names like “Claude Opus”. That mislabels the model in the UI, in logs and in your records of processing.
• Ollama cloud models are not local. Once you switch to a cloud model, prompts leave the machine for Ollama’s service, a processor you would need to assess separately.
• Cowork assumes Claude-grade tool use. Long agentic tasks with file operations depend on reliable tool calling and large context. Small local models may fail silently in the middle of a task. We have not benchmarked this and found no vendor data on it.

If a local model is the goal because data must not leave the building, a self-hosted gateway in front of Bedrock or Google Cloud in an EU region is the supported path.

What should IT and data protection check before sign-off?

Sign-off needs six answers: which provider and region, which processor contract, where local history lives and how long, which Anthropic egress remains, which extra processors (search, MCP servers) are added, and how you audit usage without the Compliance API. The checklist below maps each to a concrete configuration key.

Check	Where to verify	Key or setting
EU inference only	Bedrock profile IDs or Google eu region	inferenceModels, inferenceVertexRegion
Processor and DPA	AWS or Google contract; app under Anthropic Commercial Terms	Legal page
Local history retention	Device encryption, retention period, legal hold	coworkSessionRetentionDays, sessionRetentionHold
Anthropic egress	Telemetry, updates, previews	disableEssentialTelemetry, disableNonessentialTelemetry, disableAutoUpdates
Agent network access	Hosts the sandbox may reach	coworkEgressAllowedHosts, disabledBuiltinTools
Additional processors	Web search vendor, remote MCP servers	managedMcpServers
Audit trail	Own OpenTelemetry collector	otlpEndpoint
Prompt injection	Anthropic: risk is “non-zero”	Restrict folders and egress

Local history sits in ~/Library/Application Support/Claude-3p/ on macOS and %LOCALAPPDATA%\Claude-3p\ on Windows. By default it is kept until the user deletes it, so a retention period belongs in the first profile you ship. Each Cowork session also writes an HMAC-chained audit.jsonl, which helps with incident review. For the broader GDPR comparison across providers, see our Claude GDPR provider comparison.

What does Claude Cowork cost in the enterprise?

On 3P there is no seat fee: you pay tokens at your cloud provider’s rates. With the Bedrock EU profile, Claude Sonnet 5 costs 2.20 USD input and 11 USD output per million tokens, Opus 5 costs 5.50 and 27.50 USD. Claude Enterprise instead charges 20 USD per seat per month plus usage at API rates. All seat plans are compared in our Claude pricing plans overview.

Option	Seat price	Token price (Sonnet 5, in/out per MTok)	EU inference
Cowork on 3P, Bedrock EU profile	none	2.20 / 11 USD	Yes
Cowork on 3P, Google eu multi-region	none	2.20 / 11 USD	Yes
Claude Enterprise (1P)	20 USD/seat/month, annual	API rates (2 / 10 USD)	No EU option documented
Claude Team Standard (1P)	20 USD/seat/month annual, 25 monthly	included in plan limits	No EU option documented

Prices come from Amazon Bedrock pricing, the Google Cloud price list and claude.com/pricing; all are USD before tax. A worked example with assumed volumes: a user who processes 2 million input and 200,000 output tokens a day on Sonnet 5 costs 4.40 + 2.20 = 6.60 USD per day on the EU profile. Prompt caching lowers the input share sharply, which is why Anthropic warns that gateways must not strip cache_control. Measure real consumption in a pilot before budgeting; the 3P cost logic is compared in detail in Claude Enterprise vs. Bedrock, and more numbers are on our statistics page.

FAQ

Is Claude Cowork GDPR compliant on 3P?

The mode gives you the building blocks: EU inference via Bedrock or Google Cloud, local storage and configurable retention. Compliance still depends on your contract with the cloud provider, your records of processing, device security and any extra processors such as search vendors. No product is GDPR compliant on its own.

What does Claude Cowork on 3P cost per user?

There is no seat licence. You pay the cloud provider per token, for example 2.20 USD input and 11 USD output per million tokens for Sonnet 5 on the Bedrock EU profile. Real cost per user depends on task volume and cache hit rate, so run a pilot and read the Bedrock or Google billing data.

Claude Enterprise vs. Cowork on 3P: which should we choose?

Choose Claude Enterprise when first-party processing by Anthropic is acceptable: you get mobile, web access, sharing, the Compliance API and faster features. Choose 3P when data must stay in an EU region under your cloud contract, or when you want token billing without seats.

Does Anthropic see our prompts in Cowork 3P mode?

With Bedrock or Google Cloud, Anthropic’s docs say the app sends prompts, responses and files only to your configured endpoint and to local disk. Telemetry to Anthropic excludes content and can be disabled. With Microsoft Foundry, Anthropic operates the models and processes content as an independent processor for Microsoft.

Can Cowork use Claude Fable 5.1 in the EU?

Not on Bedrock with EU residency: Fable 5.1 has no EU inference profile and is reachable in Europe only via the global endpoint. Google Cloud lists Fable 5.1 on the EU multi-region, with prompts stored up to 30 days for abuse monitoring. Check whether that retention fits your policy.

Where is Cowork chat history stored in 3P mode?

On the user’s device: ~/Library/Application Support/Claude-3p/ on macOS and %LOCALAPPDATA%\Claude-3p\ on Windows. It stays until the user deletes it unless you set coworkSessionRetentionDays or chatSessionRetentionDays (1 to 3650 days). Include those folders in your device encryption and backup policies.

Sources

1. Claude docs: Claude Desktop on 3P overview (19 September 2026)
2. Claude docs: Features on 3P (19 September 2026)
3. Claude docs: Deploy Claude Desktop on 3P with Amazon Bedrock (19 September 2026)
4. Claude docs: Deploy Claude Desktop on 3P with Google Cloud's Agent Platform (19 September 2026)
5. Claude docs: Deploy Claude Desktop on 3P with an LLM gateway (19 September 2026)
6. Claude docs: Configuration reference (19 September 2026)
7. Claude docs: Telemetry and egress (19 September 2026)
8. Claude docs: Installation and setup (19 September 2026)
9. Claude docs: Legal and compliance (3P) (19 September 2026)
10. Claude Help Center: Use Claude Cowork on Team and Enterprise plans (19 September 2026)
11. Ollama docs: Claude Desktop integration (19 September 2026)
12. AWS Bedrock docs: Claude Sonnet 5 model card (18 September 2026)
13. Amazon Bedrock pricing (18 September 2026)
14. Claude pricing (19 September 2026)