Anthropic Ships Self-Hosted Claude Agents for Enterprise

Anthropic opened private beta access for a self-hosted edition of Claude Managed Agents on November 12, 2026, letting enterprises run the entire agent infrastructure inside their own cloud or on-premise environment. The move shifts operational control from Anthropic to the customer, with credentials, session data, and worker sandboxes all staying behind the enterprise firewall while model calls route to external providers or internal gateways.
Access is gated through Anthropic account teams, with pricing determined by quote. The release lands the same day Google opened its Agent Gateway preview for Kubernetes clusters, marking a parallel push toward enterprise-controlled agent deployment infrastructure.
Architecture: Customer-Operated Infrastructure
In the self-hosted configuration, customers operate three core components: the Managed Agents server, a PostgreSQL database for session history, and the sandboxed workers where agents execute shell and file operations. The Claude model itself remains external, so prompts and completions still traverse the network, but customers control where those requests go. Model calls can be routed to Anthropic directly, to Amazon Bedrock, to Google's Agent Platform, to Microsoft Foundry, or to an internal gateway that sits behind the corporate firewall.
This architecture splits control from computation. Anthropic no longer holds the agent harness or session state, but the underlying Claude model continues to run outside the customer perimeter. Enterprises gain custody of execution logs, credentials, and runtime state without forking the model weights or running local inference.
Credential Isolation and Security Model
Credentials never enter the sandbox where agent code runs. For MCP tool calls and supported API integrations, a proxy sitting outside the sandbox injects the credential at the moment of the request. Self-hosted customers can store those secrets in their own secrets manager, and the harness that calls Claude is kept separate from the sandbox that executes tool code. Agents are restricted to allowlisted hosts, preventing arbitrary network traversal.
Every agent step is logged and exported over OpenTelemetry, letting security teams route telemetry into their existing monitoring stack. According to Help Net Security, this gives enterprises full visibility into agent behavior without relying on Anthropic's observability layer.
Operational Trade-Offs
Customers pick up the operational burden in return for control. Patching, firewall rules, database backups, and configuration management all become the enterprise's responsibility. The hosted version offloads that work to Anthropic, but the self-hosted edition hands it back. For organizations with strict data residency requirements or regulatory constraints that prohibit sending session state to third-party infrastructure, the trade-off is deliberate.
The article does not name any beta customers or disclose pricing structure beyond confirming that access is by quote. Anthropic account teams are handling onboarding, suggesting the beta is targeted at large enterprises rather than open to the general developer pool.
Release Date and Availability
Officially launched on November 12, 2026. Access is currently restricted to private beta participants, with onboarding managed through Anthropic's enterprise sales organization. No timeline for general availability has been announced.
What This Means
The self-hosted Claude Managed Agents beta represents a structural shift in how frontier labs are packaging agent capabilities for enterprise deployment. By letting customers run the agent harness inside their own perimeter while still relying on external model inference, Anthropic is splitting the deployment model into a hybrid architecture that satisfies data residency requirements without requiring on-premise model hosting. This approach competes directly with fully self-contained solutions and cloud-native agent platforms, carving out a middle ground where the model stays external but the execution environment does not. Enterprises now have the option to keep credentials, session state, and telemetry entirely within their own control plane, a requirement that has blocked adoption in regulated industries. The same-day timing with Google's Agent Gateway preview suggests the enterprise agent infrastructure category is maturing rapidly, with multiple vendors converging on similar hybrid-control architectures.
on Emergent today





