Self-Hosted Claude Managed Agents Explained (November 2026)
The short answer
On November 12, 2026 Anthropic opened a private beta of self-hosted Claude Managed Agents: the same agent platform, but every part runs inside your own cloud account or data centre. You operate the Managed Agents server, its database and the workers where agents run shell and file tools; model requests go wherever you point them. Access is through your Anthropic account team.
What runs where
| Option | Sandboxes | Service (API, agent loop, session history) | Model |
|---|---|---|---|
| Cloud sandboxes (default) | Anthropic | Anthropic | Anthropic |
| Self-hosted sandboxes | You | Anthropic | Anthropic |
| Self-hosted Managed Agents (private beta) | You | You | Anthropic, or a provider you choose |
A self-hosted deployment has four parts:
- Server (
cma-server) — serves the API, runs the agent loop and stores agents, sessions and history. It never runs the agent’s tools. - Database — PostgreSQL 16 or later.
- Workers —
ant beta:workerprocesses or SDK workers that poll the server and run tools in your sandboxes. Anthropic provides a Docker Compose stack and acma-sandboxesHelm chart that start each session’s worker in its own container or pod. - Console and admin CLI — a web console on your server behind your identity provider, and
cma-server-adminfor environments and keys.
Why companies want it
Anthropic gives two reasons to choose it over self-hosted sandboxes: storage on your infrastructure (agents, sessions and history in your database) and your choice of inference provider — Amazon Bedrock, Claude models on Google Cloud, Microsoft Foundry or your own model gateway. If your organisation has zero data retention with Anthropic, it applies to the model requests as for any Messages API call. The trade-off: you secure the server, install each release yourself, and some features exist only on the hosted service.
The security model
Anthropic built Managed Agents around four principles that hold in every hosting option:
- Brain and hands are split — the harness that calls Claude runs outside the sandbox, so session history and secrets are not in the box where untrusted code runs.
- Credentials stay out of the sandbox — vaults use a proxy outside the sandbox to add credentials to MCP and supported API calls; self-hosted customers can use their own secrets manager.
- Network access can be limited to a list of allowed hosts — for a build-fixing agent, perhaps only the git host and package registry.
- Every session leaves a full record, exported over OpenTelemetry.
Anthropic’s docs warn that the worker’s bash tool is a real shell with the worker’s permissions; isolation comes only from the container or VM it runs in, and each trust level needs its own deployment.
Should you wait for it?
Only if a regulator, customer contract or internal policy requires that agent session data stays in your environment, or that model traffic goes through your cloud provider or gateway. Otherwise the hosted service or self-hosted sandboxes are less work. If you are comparing platforms, see Microsoft Foundry vs AWS AgentCore vs Claude Managed Agents and our guide to agent sandbox platforms for network and secrets control.
Last verified: November 16, 2026.