Self-hosted Claude Managed Agents: Anthropic's agent service now runs on your infrastructure
Anthropic opened self-hosted Claude Managed Agents in private beta: you host the agent service, its database, and the sandboxes, while model calls go to the Claude API or a provider you choose. What runs where, which credentials exist, what still leaves your network, and what to check before committing.
Anthropic has opened self-hosted Claude Managed Agents in private beta. Until now, the Managed Agents service ran at Anthropic: you defined an agent, started a session, and Anthropic hosted the agent loop, the session storage, and the sandboxes where tools run. The new option moves all of that except the model onto your own infrastructure, whether that is a cloud account or an on-premise data center. Access goes through your Anthropic account team; there is no self-serve download yet.
The short version for builders: this is the first time the full agent service, not just the sandboxes, can live inside your perimeter. If you have been blocked on session history leaving your cloud, this is the release that removes that argument.
What self-hosted actually means
Managed Agents has two hostable parts, and the naming now tells you which one you hold:
- Cloud sandboxes (default): Anthropic hosts the sandboxes, the service, and the model.
- Self-hosted sandboxes: you host the sandboxes; Anthropic still runs the service and the model.
- Self-hosted Managed Agents: you host the sandboxes and the service; the model runs at Anthropic or at a provider you choose.
The model options on the self-hosted tier are the Claude API, Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry, or your own model gateway. Both the hosted service and your server expose the same API, so the SDK and the ant CLI work unchanged against your deployment, and the rest of the Managed Agents documentation applies.
When does Anthropic say to choose it? Two triggers: you must store agents, sessions, and session history on your own infrastructure, or model calls must go to a specific inference provider. If you only need tools to run on your machines, self-hosted sandboxes remain the lighter option.
What runs where
One deployment is one or more servers, a database, and your sandboxes:
- Server (cma-server): serves the API, runs the agent loop, and stores agents, sessions, and session history. It calls your inference provider when the agent needs the model, and never runs an agent's shell or file tools itself.
- Database: PostgreSQL 16 or later. Skip it and the server keeps data in memory; a restart erases everything.
- Workers: programs that take work from the server and run the agent's shell and file tools in your sandboxes, via an ant beta:worker command or an SDK worker.
- Console on your server: a webpage at the server's address where people sign in through your identity provider.
- Admin CLI (cma-server-admin): creates environments and their keys; runs on Linux (x86-64 and Arm) and macOS on Apple silicon.
The three credentials
The security model splits across three keys, and it is worth reading before any design review:
- Server API key: held by your applications, the ant CLI, and the admin CLI. Authenticates requests to your server. Plays the role a Claude API key plays against the hosted service.
- Inference credential: held by the server, for its model calls. Can be a Claude API key, a cloud provider credential, or your gateway token. Applications and workers never see it.
- Environment key: held by workers that poll the server. Scopes access to one environment's work queue and sessions.
What leaves your network
This is the section to read twice. For every model call, the server sends the full conversation so far, including prompts, tool inputs, and tool results, to your inference provider. Self-hosting moves storage and execution to you; it does not keep prompts inside your network unless your inference provider is too.
Two more details from the data-and-access page:
- When the Claude API is the provider, each call carries up to three headers with your server-generated session, agent, and environment IDs. Anthropic says these let it attribute usage, that the IDs hold no content, and that no setting turns them off. The IDs do encode creation time.
- Otherwise the server connects only to a fixed list: your inference provider, the database, credential-issuing services, a model-listing service, the MCP servers and Git hosts you allow, an OpenTelemetry collector if configured, your SSO identity providers, and a secret manager or KMS if configured. Web search and web fetch tools run at the inference provider and stay off until you enable them.
Anthropic states it has no remote access to the server or its logs. Their suggestion for verification is blunt: compare your egress proxy or firewall log against that connection list.
Your security responsibilities
The agent's bash tool is a real shell with the worker process's permissions and network access; the worker does not isolate it. Isolation comes from what the worker runs in. The Docker Compose stack and the cma-sandboxes Helm chart start each session's worker in its own container or pod. If you start a worker yourself, Anthropic's guidance is to run it in a disposable container or VM holding nothing you need to protect, and each trust level needs its own deployment.
Everything the hosted security model assigns to Anthropic is now yours: securing the server, hosting memory stores, detecting unusual usage.
Getting started: Compose for dev, Helm for production
The quickstart runs the whole stack on one machine with Docker Compose: server, PostgreSQL, and workers that give each session its own sandbox container. It is explicitly for development, serving plain HTTP on port 8400 bound to loopback, with a pointed caveat that Docker Engine releases before 28.3.3 can expose that port to the local network. Production means two Helm charts on Kubernetes. Prerequisites are the release archive (via your account team), a Claude API key, Docker with the Compose plugin, at least two CPUs on x86-64 or Arm, openssl, and the ant CLI.
Each release archive ships an operator-guide.md for that version, and Anthropic suggests having Claude read it to help configure your server. The server software is licensed under a dedicated Claude Managed Agents Server agreement, not an open source license.
What I would check
My untested suggestions before treating this as a deployment target:
- Run the Compose quickstart on a throwaway machine and watch actual egress against the documented connection list before any architecture decision.
- Decide your inference provider first; it determines what conversation data leaves your network regardless of where the service runs.
- Read the differences-from-hosted page before committing; parts of it sit behind a login, so request full access through your account team early. Some features exist only on the hosted service.
- Plan the per-trust-level deployment split up front; Anthropic is explicit that one deployment is one trust boundary.
- Check what the license agreement says about your use case before legal review becomes the blocker.
What I would not conclude
Do not conclude self-hosted means fully air-gapped; conversation content still flows to whichever inference provider answers model calls, and the attribution headers to the Claude API cannot be disabled. Do not assume feature parity with the hosted service. Do not treat the Compose quickstart as a production pattern. And do not read private beta as generally available; access today runs through an Anthropic account team.
Frequently asked questions
What is self-hosted Claude Managed Agents?
It is Anthropic's Managed Agents service running on infrastructure you control. You host the server, its PostgreSQL database, and the sandboxes where agent tools execute. Model calls go to the Claude API or an inference provider you choose, such as Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry, or your own gateway.
How is it different from self-hosted sandboxes?
Self-hosted sandboxes move only the tool-execution layer to you while Anthropic still runs the service and stores sessions. Self-hosted Managed Agents moves the service and its storage too, so agents, sessions, and session history live in your database.
Does any data still leave my network?
Yes. Every model call carries the full conversation so far, including tool inputs and results, to your inference provider. When that provider is the Claude API, calls also carry session, agent, and environment ID headers that cannot be turned off. Beyond that, the server connects only to destinations you configure or that the documented fixed list names.
Who can get access right now?
It is a private beta. Anthropic's docs direct you to contact your Anthropic account team to request access to the release archive.
What do I need to run it locally?
The release archive, a Claude API key, Docker with the Compose plugin, a host with at least two CPUs on x86-64 or Arm, openssl, and the ant CLI. The Compose stack is for development only; production deployments use two Helm charts on Kubernetes.
Is the server software open source?
No. It is distributed as a release archive under a dedicated Software License Agreement for Claude Managed Agents Server.
Does self-hosted support the same API as the hosted service?
Yes. Anthropic says both expose the same API, so the SDK and the ant CLI work against your server, with exceptions listed on a differences-from-hosted page that is partially gated behind a login.
Sources
- Self-hosted Claude Managed Agents overview | Claude Platform Docs - read November 14, 2026. Hosting options, when to choose it, components, the three credentials, security responsibilities, and the license note.
- Data and access in self-hosted Claude Managed Agents | Claude Platform Docs - read November 14, 2026. The full connection list, the attribution headers, and what model calls carry.
- Get started with self-hosted Claude Managed Agents | Claude Platform Docs - read November 14, 2026. Compose quickstart, prerequisites, the port 8400 caveat, and the Helm production path.
- TLDR AI newsletter, November 13, 2026 - the item that surfaced the beta.
Not read in full: the differences-from-hosted page (partially behind login) and the release's operator-guide.md (inside the access-gated archive). Launch timing and pricing are not stated in the public docs; coverage of the beta began November 12 per secondary reports. I did not run the software.
Related reading
- Anthropic's always-on agents guide: the five primitives behind agents that never sit idle
Anthropic published a guide on building always-on agents: stateful, proactive agents that manage context across sessions. The core idea is that the agent is not actually running all the time - it sleeps between runs, and five architectural primitives decide when it wakes, what it remembers, and what reaches you.
- Anthropic's Five Claude Exploit Chains: What the Frontier Red Team Post Claims, and What It Leaves Out
Anthropic says an unreleased internal model found and chained vulnerabilities in Chrome, Safari, Android, Windows and Linux. What the October 27 post claims, the disclosure terms, the defender advice, and the numbers it does not give.
- AI Model Behavioral Consistency: What the Research Shows
Anthropic measured values across 300,000 conversations, 3 model versions, and 20 languages. The same AI behaves differently depending on who asks and how.
The build log
Follow the work.
New essays and prototypes, in your feed reader. Email subscriptions are paused while I connect a working delivery service. No email is collected here.
Follow via RSS