Claude Managed Agents can now run entirely on your own servers
A new self-hosted mode lets you keep agent sessions, history and credentials on your own infrastructure, and route model calls to Bedrock, Google Cloud or your own gateway.
Anthropic has opened a private beta for self-hosted Claude Managed Agents, letting customers run the full agent service, not just the sandboxes, on their own cloud account or on-premise data centre. Previously, self-hosting only covered the sandboxes where an agent’s tools run; the service itself, including the API, the agent loop and all session storage, always stayed with Anthropic.
The self-hosted server (cma-server) needs PostgreSQL 16 or later, and deploys via Docker Compose for a single machine or two Helm charts for Kubernetes. Model calls can go to Anthropic’s own API, or to Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry, or a model gateway of your choosing. It exposes the same API as the hosted service, so the existing SDK and ant CLI work against it unchanged. Access requires contacting your Anthropic account team.
Why it matters: if a compliance requirement has been the thing stopping you building on Claude’s agent platform, specifically needing session history and credentials to never leave your own network, there’s now a documented path to do that without losing the SDK and tooling.
The caveat: this is a private beta behind an account-team request, not a self-serve signup, and you take on running and securing the server, its database and your sandboxes yourself.