Step by step, the server-side version of Claude Managed Agents is taking shape.
This beta offering provides the tooling to run Claude as a self-contained agent. Access is enabled by default for all API accounts.
We could already self-host the sandbox environments. We can now do the same with the entire orchestration server.
Anthropic offers a Compose stack for development and Helm charts for production. The worker software is the same as with the self-hosted sandbox environments. It is possible to choose the inference provider.
A self-hosted mode for Claude Code as well… outside the control plane
Claude Code also offers a self-hosted mode, available in public beta for Team and Enterprise subscriptions. Disabled by default, it encompasses the runners and, potentially, an autoscaling orchestrator. It can be configured to route inference requests to Amazon Bedrock or Google Cloud Agent Platform.
Beyond compliance benefits, this model enables access to internal services without exposing them to the Internet. Or to preload compilers, SDKs, and the CLI inside the images of the runners. The control plane (the Claude.ai interface, queuing, and session management) remains with Anthropic.