Anthropic opened a private beta Nov. 12 that lets companies run Claude Managed Agents in their own environments, extending control over data, credentials and network access, according to its announcement.
The self-hosted package includes the agent harness, which coordinates work, and the sandboxes where agents use tools. Customers can also operate the supporting servers, database and workers for shell and file actions. They can keep session history in their own database and credentials in a secrets manager.
Anthropic said the system keeps the agent’s work coordinator separate from the sandbox that runs tools. Organizations can set a list of permitted network destinations and export activity data through OpenTelemetry for monitoring. The sandbox does not contain session history or vault credentials.
Model requests can go to Anthropic, Amazon Bedrock, Google’s Agent Platform, Microsoft Foundry or an internal gateway. Teams can keep using the same integration points when they move an agent to the self-hosted setup, Anthropic said.
Anthropic said organizations with a zero data retention agreement with the company can apply the same terms to Claude requests as to other Messages API calls.
Anthropic previously ran Managed Agents as a hosted service, then introduced customer-run sandboxes while retaining the rest of the infrastructure; the beta moves more components into customer operations. Anthropic cites repairing failed builds, checking releases, sorting alerts and handling bug backlogs as possible tasks. Organizations can request access by contacting their Anthropic account team.