Anthropic has opened a private beta of self-hosted Claude Managed Agents, a deployment option for teams that want the agent service itself to run in their cloud or on-premises environment. The announcement is about where the agent platform and its operational data run; it does not say that a Claude model must run locally. Developers can choose where the model requests go, which may still be an external provider.
What moves into the developer's environment?
Anthropic says the self-hosted version runs the Managed Agents harness, sandboxes, credential handling, session history, database and workers in the customer's environment. A worker executes an agent's shell and file tools. Developers use the same core APIs as the hosted service, while operating the service and its underlying systems themselves. Anthropic says the beta is available by request through an account team.
That is a broader option than self-hosting only a sandbox. With a self-hosted sandbox, the worker and tool execution can remain on the customer's infrastructure, while the Managed Agents service stays operated by Anthropic. With the new beta, the service control plane and session store can also be hosted by the customer. These options answer different operational needs and should not be described as the same architecture.
Where can model requests go?
Anthropic lists direct requests to Anthropic, Amazon Bedrock, Google's Agent Platform, Microsoft Foundry, or the customer's own gateway. The route depends on the provider and configuration selected by the team. Sending a request to Anthropic or another cloud provider means model input is still sent to that provider; self-hosting the agent service does not by itself keep all inference data inside the customer's network. A gateway can control routing, but its actual model destination and logging need separate review.
Anthropic describes a split between the harness that decides which tool to call and the sandbox where the agent's code runs. It says credentials can stay outside the sandbox and be attached through a proxy, network access can be restricted to allowed hosts, and each session has an event record that can be exported through OpenTelemetry. These are vendor-described design features. A team using the private beta still operates the server, database, workers, network rules, secrets and monitoring in its environment.
Anthropic product announcement: service components and model-request destinations
Claude Platform quickstart: development Compose stack and production deployment guidance
Claude Platform documentation: what a self-hosted sandbox does and does not move
A beta deployment still needs production decisions
Anthropic's quickstart labels its Docker Compose setup for development rather than production. It serves plain HTTP and binds the default port to the Docker host's loopback address; the docs also warn that some older Docker Engine releases can expose that port to the local network. Anthropic recommends its release Helm charts for production. This is a reminder to follow the deployment documentation for the exact release instead of treating a working local stack as a hardened service.
Teams comparing deployment options can start with four questions: which records stay in the environment they control, which tool outputs are sent to the model provider, who operates the worker and its network permissions, and how the organization will review session traces. Self-hosting can change custody and control, but it also moves infrastructure and incident-response duties to the operator.
For a maker, the release is most relevant when an agent needs access to internal code or systems and the organization needs tighter control over the service running it. It is not a shortcut to local-only inference or a substitute for deciding which actions an unattended agent is allowed to perform.