Anthropic has introduced a self-hosted deployment option for its Claude Managed Agents platform, allowing enterprise customers to bring the agent harness, sandboxes, session history, and credential management within their own cloud or on-premise infrastructure. Previously, while customers could host the isolated sandbox environments where agent code executes, Anthropic maintained control over credential handling and session logging. The new self-hosted architecture, which entered private beta on November 12, 2026, shifts the operational responsibility for the Managed Agents server and its associated PostgreSQL database to the customer.
Under this new deployment model, the customer is responsible for maintaining the Managed Agents server, the database containing session history, and the worker nodes that execute shell and file tools. While the core model processing still occurs externally—meaning prompts and completions are transmitted to and from Anthropic or other supported platforms like Amazon Bedrock, Google’s Agent Platform, or Microsoft Foundry—the customer now retains control over the underlying infrastructure. For organizations with existing zero data retention agreements, these model requests are handled according to the same standards as standard Messages API calls.
A central component of this self-hosted offering is the management of credentials. Anthropic’s security framework for Managed Agents dictates that credentials remain separated from the sandbox environment where the agent’s code runs. Instead, a proxy mechanism outside the sandbox attaches necessary credentials to Model Context Protocol (MCP) tools and specific supported API calls. By moving to a self-hosted model, customers can integrate these credential flows directly with their own internal secrets management systems, ensuring that sensitive access tokens do not reside within the sandbox or the vendor's managed environment.
Anthropic emphasizes that this protection is currently limited to MCP tools and a defined set of supported API calls. Consequently, security teams are advised to audit any additional services or external endpoints that an agent might interact with before deploying them into production environments. The responsibility for ensuring that these secondary services are secure rests with the customer, as does the task of maintaining and patching the self-hosted server and database infrastructure.
To assist with oversight, the platform includes granular control over network traffic through allowlists. Customers can restrict an agent’s network access to a specific set of authorized hosts, such as internal git repositories or package registries, by enforcing these limits via their own firewall configurations. This configuration is intended to prevent agents from communicating with unauthorized or malicious external endpoints.
Transparency and auditability are maintained through comprehensive logging. Every action taken by an agent, including individual tool calls and the corresponding responses, is recorded. These logs are designed to be exported via OpenTelemetry, enabling security teams to ingest agent activity data into existing monitoring and observability stacks. This allows for the correlation of agent behavior with other system events, providing a centralized view of activity within the customer’s perimeter.
This expansion of the Managed Agents platform is aimed at organizations utilizing agents for unattended tasks, such as automated build remediation, release testing, alert triaging, and bug backlog management. By moving the harness and session history behind the corporate firewall, Anthropic intends to provide teams with greater control over the lifecycle and security posture of their autonomous agents. Access to the private beta is currently being managed through direct coordination with Anthropic account teams.





