Anthropic opened a private beta of self-hosted Claude Managed Agents on Nov. 12, Help Net Security reports. Companies can now run the agent harness, sandboxes, credentials and session history in their own cloud or on-prem. You request access through your Anthropic account team.
- You run three pieces: the Managed Agents server, its PostgreSQL database (where session history lives) and the workers where agents run shell and file tools. The APIs are the same as the hosted service.
- Model calls can go to Anthropic, Bedrock, Google’s Agent Platform, Microsoft Foundry or your own gateway. Claude itself still runs outside your network.
- The harness and sandbox are kept apart. Secrets stay outside the sandbox, and a proxy attaches credentials to MCP and API calls. Self-hosters can keep those in their own secrets manager.
My take
Keeping credentials out of the sandbox is the right design. It’s the same lesson as the Darktrace Claude Code research in this batch. Self-hosting also means you own that PostgreSQL database, and it just got 52 CVE fixes today. Patch it.
God Bless, and Happy hacking!
