Playbooks · Teams
Run Claude Managed Agents in your own cloud
Written November 12, 2026
Anthropic's private beta lets you run the whole Managed Agents service — API, agent loop, sessions — in your own cloud or on-prem.
What changed
Managed Agents started as a hosted service, and Anthropic later added self-hosted sandboxes, which moved only tool execution into the customer's environment while Anthropic kept the service itself. The new option moves the rest. With self-hosted Managed Agents, you host the sandboxes and the service: the API, the agent loop, and the storage for agents, sessions, and session history. Anthropic's announcement puts the same point in operational terms: you operate the Managed Agents server, its database, and the workers where agents run their shell and file tools.
The audience is teams whose agents already work unattended on internal systems — fixing broken builds, testing releases, triaging alerts, working through a bug backlog — and who need session data and credentials to stay inside their own perimeter for that work.
What you run
- The server (
cma-server). It serves the API, runs the agent loop, and stores agents, sessions, and each session's history. It calls your inference provider when the agent needs the model. The server never runs an agent's shell or file tools. - The database. A persistent deployment needs PostgreSQL 16 or later. Without a database, the server keeps its data in memory, and a restart erases it.
- The workers. Programs that take work from the server and run the agent's shell and file tools in your sandboxes: an
ant beta:workercommand or an SDK worker. - The console and the admin CLI. A web console at the server's address, where people sign in through your identity provider to create agents and run sessions, and
cma-server-admin, which creates environments and their environment keys on Linux and macOS on Apple silicon.
Your own applications reach the server with the SDK or the ant CLI. Both have the same API as the hosted service, so existing code connects unchanged, and the rest of the Managed Agents documentation applies. Three credentials divide the access: a server API key for your applications and CLIs, an inference credential the server holds for its model calls, and an environment key for each environment's workers.
Where the model calls go
Hosting the service does not host the model. By default, model calls go to the Claude API. You can point them instead at Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry, or your own model gateway. For each call, the server sends the conversation so far, including tool inputs and tool results, to that provider. A call to the Claude API also carries the IDs of its session and of that session's agent and environment. If your organization has a zero data retention agreement with Anthropic, it applies to the requests sent to Claude in the same way as any other Messages API call.
The security responsibilities move with the service. Securing the server, hosting memory stores, and detecting unusual usage are yours, and each trust level needs its own deployment. Credentials can stay in your own secrets manager, network access can be limited to a list of allowed hosts per environment, and every session leaves a full record of tool calls and results that exports over OpenTelemetry.
How to get in
This is a private beta, not a public download. To request access to the download, contact your Anthropic account team. You install each release yourself, and each release's archive includes an operator guide for that version. Plan the operating work before you ask: a server, a PostgreSQL database, sandboxes, and upgrades are now part of your platform, not Anthropic's.
Limits
- Private beta: access runs through an Anthropic account team, and there is no download link to promise a team today.
- Some features exist only on the hosted service. The documentation lists the differences, and a team that depends on a hosted-only feature has to weigh that before moving.
- The model still runs outside your network. Prompts and tool results leave for the inference provider you choose on every model call; what stays home is the service, the sessions, and the credentials.
Related playbooks
- Kimi Hosted Agents: a managed harness in the cloud
- Verify what managed settings auto-apply after 2.1.316
- Run work that spans several repos in one Copilot Project, and keep computer use on a short leash
- Anthropic's model-protection rule is now in force
Sources
- Operate Claude Managed Agents inside your security perimeter — Claude by Anthropic
- Self-hosted Claude Managed Agents — Claude Platform Docs
Back to: Teams playbooks hub · Also: Claude Meetings, Sheets and Whiteboards land · Lazarus: ten neurons can backdoor an agent