November 12, 2026. Anthropic announced a private beta of self-hosted Claude Managed Agents today. The important change is ownership of the agent service, not a downloadable Claude model. Organizations can run the service and its execution environments on their infrastructure, while choosing the provider that answers model requests. Access requires contacting an Anthropic account team. Announcement.
Two boundaries to review
The existing self-hosted sandbox option moved tool execution to the customer. This option also moves the service and session storage. That matters for IT teams responsible for retention, incident response and internal-system access. It also creates an operating responsibility: someone must maintain the server, database and workers, and install updates. Deployment overview.
Self-hosting does not make the conversation stay entirely inside the organization. The data-access documentation says model calls include prompts, tool inputs and tool results. Worker network access is a separate boundary from the server’s connections. Map both before approving a deployment; review the chosen inference provider’s retention terms rather than assuming local storage settles the question. Data and access.
Do not treat migration as a server-address change
The documented differences include no automatic transfer of existing sessions, no hosted webhooks or scheduled deployments, and no built-in refresh flow for MCP OAuth credentials. A workflow that relies on those features needs an explicit replacement, not just a successful first chat. Differences from the hosted service.
A useful first trial
Start with non-sensitive sample documents and a read-only internal tool. Record which system receives each prompt and tool response; check that a denied destination really fails; and assign owners for patching, credential renewal and incident review. Treat these as proposed acceptance checks, not tests we have performed.
This is a private-beta deployment option, not proof of better security or lower cost. The decision is whether the extra control solves a real requirement and whether the team can operate it safely.