Self-hosted Claude Managed Agents: what you run, what still leaves your network
Anthropic has opened a private beta for self-hosted Claude Managed Agents, enabling the entire service, including the server, its PostgreSQL database for session history, and workers for shell and file tools, to run within a user's cloud or on-premise infrastructure. Access is available by request through Anthropic account teams.
Previously, organizations could self-host only the sandbox where an AI agent's code executed, while Anthropic managed credentials and agent activity records. The new private beta significantly expands this control, allowing users to operate the Managed Agents server, its PostgreSQL database which stores session history, and the workers that run shell and file tools directly within their own network.
This shift means that critical components such as the agent loop, session history, and vault secrets now reside within the user's domain. Organizations can also manage credentials in their own secrets manager, with a proxy attaching them for specific tool and API calls. Security is further bolstered by per-environment host allowlists and the ability to export all tool call logs and results via OpenTelemetry.
However, a key distinction remains: the Claude large language model itself continues to operate externally. Model requests, consisting of prompts sent out and completions returned, are directed to providers like Anthropic, Amazon Bedrock, Google's Agent Platform, Microsoft Foundry, or a user's own gateway. These interactions are covered by zero data retention agreements, but the core AI processing does not occur within the self-hosted infrastructure.
The trade-off for this increased operational control is the responsibility for infrastructure management. Users are now accountable for patching, securing, and configuring the server, database, and firewall. The Next Web also reports that no migration path exists from the hosted service, and features such as webhooks and scheduled runs are currently absent from the self-hosted beta.
Share this article
Related reading
6 stories
Claude Code is quietly becoming a hardware hacker's best helper, and Linux support is next
We recently explored how Claude Code is becoming a key helper for hardware hackers, with Linux support on the horizon.

Vietnamese Travellers Can Now Scan NETS QR Codes to Pay in Singapore

Weixin Pay Sees 84% Rise in Overseas E-Wallet Transactions

BPC Pairs SmartVista with Huawei for Payment Modernisation

