NewsAI

Anthropic launches self-hosted Claude Managed Agents for enterprise control

Anthropic has released a self-hosted version of its Claude Managed Agents service, allowing companies to run agent infrastructure on their own servers while keeping model calls external.

3 min read

SCROLL

Self-hosted option launches

Anthropic has introduced a self-hosted version of its Claude Managed Agents service, enabling enterprises to deploy the agent infrastructure within their own cloud accounts or data centers. According to release notes from the Claude Platform, the service launched on November 12, 2026. Access is currently available through a private beta, with customers required to request entry through their Anthropic account team.

This update allows organizations to run the entire Managed Agents service, rather than just isolated sandboxes, on their own hardware. Previously, customers could host the sandbox environment where agent code executes, but Anthropic managed the broader service components, including credential handling and session history storage. With the new self-hosted option, companies take full control of the agent harness, the PostgreSQL database that stores session history, and the workers that execute shell and file tools.

How the architecture works

In the self-hosted configuration, the customer is responsible for operating three main components: the Managed Agents server, the database, and the worker nodes. Session histories and credentials remain on the customer's systems, providing greater data sovereignty. However, the underlying Claude model does not run inside the customer's network. Each time an agent requires a model response, the request is sent to an external provider.

Customers can direct these model requests to Anthropic directly, or to third-party platforms such as Amazon Bedrock, Google Cloud, or Microsoft Foundry. Anthropic is the default provider, but organizations can choose based on their existing infrastructure agreements. For organizations with zero data retention (ZDR) agreements with Anthropic, these external model calls are covered under the same terms as standard Messages API requests.

Security and operational responsibilities

By moving the infrastructure in-house, customers assume significant operational responsibilities. They must secure, update, and patch the server, database, and firewall rules themselves. Anthropic’s security model for Managed Agents is built on four principles, which apply regardless of whether the service is hosted by Anthropic or the customer. The core design separates the agent harness, which communicates with the model, from the sandbox, where code runs. Credentials are kept outside the sandbox and are attached to requests by a proxy for supported tools and API calls.

Self-hosted customers can store these credentials in their own secrets managers. Additionally, each environment can be restricted to an allowlist of approved hosts, blocking requests to unauthorized destinations. This is particularly useful for agents performing specific tasks, such as fixing builds, where the allowlist might be limited to a git host and a package registry. All agent actions, including tool calls and responses, are logged and can be exported via OpenTelemetry, allowing security teams to integrate agent activity into their existing monitoring tools.

Limitations and migration

The self-hosted version is not a drop-in replacement for the fully hosted service. Some features available in the hosted environment, such as webhooks and scheduled runs, are currently missing in the self-hosted build. Furthermore, existing agents and data do not automatically migrate; customers must rebuild their agents and re-store their data in the new environment. Developers use the same core APIs for both hosted and self-hosted setups, but the operational burden of maintaining the underlying infrastructure shifts entirely to the enterprise.

Industry context

This move addresses growing concerns among enterprises about data residency and reliance on foreign AI providers. In Europe, regulatory and financial leaders have warned against over-dependence on external AI infrastructure, prompting companies to seek sovereign AI solutions. Competitors have adopted different strategies; for instance, OpenAI has allowed companies to run its agents within Amazon’s cloud, while Nvidia has developed hardware specifically for secure agent execution. Anthropic’s approach offers a middle ground, providing control over data and infrastructure without requiring the model itself to be deployed on-premise.