Official releaseEnterprise AI Software / Agent Runtime

Anthropic Opens Private Beta for Self-Hosted Claude Managed Agents: The Control Plane Moves Inside the Enterprise

Anthropic announced on November 12, 2026 that self-hosted Claude Managed Agents is in private beta: the harness, sandboxes, credentials, and session history run in the customer's environment, while model requests go to an endpoint the customer chooses.

Key answer

Self-hosted Claude Managed Agents is a private beta announced by Anthropic on November 12, 2026 that lets an enterprise run the whole agent service in its own cloud or on-premises, including the harness, sandboxes, credentials, and session history. Model requests still go to Anthropic, a cloud provider, or the enterprise's own gateway.

Anthropic published an article on November 12, 2026 announcing that self-hosted Claude Managed Agents is available in private beta. Managed Agents is a suite of APIs for building and deploying autonomous agents, providing the agent harness, sandboxes, credential handling, and session history. It began as a service hosted by Anthropic and later added self-hosted sandboxes. This release moves the whole service into the customer's environment.

According to the article, every part of the self-hosted service runs in the customer's own cloud or on-premises, including the harness, sandboxes, credentials, and session history. The customer operates the Managed Agents server, its database, and the workers where agents run their shell and file tools. Developers build agents with the same core APIs as the hosted service.

The customer chooses where model requests go: directly to Anthropic, to a cloud provider (the article names Amazon Bedrock, Google's Agent Platform, and Microsoft Foundry), or to the customer's own gateway. The article does not describe running the model itself inside the customer's environment; what is self-hosted is the agent control plane and execution environment. Session history is stored in the customer's own database, and an organization's zero data retention arrangement with Anthropic applies to requests sent to Claude.

The article describes four design principles and says they apply regardless of who hosts the service. First, the harness and the sandbox are split: the agent's code runs in the sandbox, while the agent loop, session history, and vault secrets stay outside it. Second, credentials stay out of the sandbox: for MCP tools and supported API calls, a proxy outside the sandbox adds the credential to the request, and the self-hosted version can also use the customer's own secrets manager.

Third, network access can be limited: each environment can be restricted to a list of allowed hosts, with all other requests blocked, and when the customer hosts the sandbox, the same limits are set in the customer's firewall. Fourth, every session leaves a full record, including tool calls and what came back, which exports over OpenTelemetry to existing monitoring tools.

Access is by request through an Anthropic account team. The article gives no pricing and no date for general availability. For enterprises evaluating agent platforms, the four principles work as a checklist for any option, including systems built in-house: whether the agent loop is separated from the execution environment, whether credentials can reach an environment that runs arbitrary code, whether outbound connections use an allowlist, and whether every step leaves an exportable record.

X Cube view

Running the agent control plane inside the enterprise is now an option offered by a model provider itself, which reflects enterprise demand to keep code, credentials, and execution records within their own boundary. Model inference in this kind of offering still happens outside or through a gateway, so enterprises should evaluate where the runtime lives and where the model lives as separate questions, and confirm their gateway can apply budgets and auditing to agent model requests.