Skip to main content

Self-hosted Claude Managed Agents: what you run and what you lose

Anthropic now lets you run the whole Claude Managed Agents service on your own infrastructure, with your choice of Bedrock, Google Cloud, Foundry or your own gateway. What you take on, which hosted features are missing, and when self-hosted sandboxes are enough.

Share -
Engineer in a white shirt at a desk of code-filled monitors pointing up at glowing server racks in a dark data centre

Self-hosted Claude Managed Agents lets you run the whole Managed Agents service, not only the sandboxes, on your own cloud account or data centre, with agents, sessions and history stored in your PostgreSQL database. Choose it when storage or the inference provider must be yours, and check the missing hosted features before you move.

What Anthropic released

Self-hosted Claude Managed Agents was announced by Anthropic on 12 November 2026 in the Claude API release notes. Until now, the furthest you could go was self-hosted sandboxes: your containers ran the agent's tools, but Anthropic ran the API, the agent loop and the storage. The new option moves that service to you as well.

Access is not self-serve. The documentation says to contact your Anthropic account team to request the download, and the server software comes under its own licence agreement. Anthropic has not published a separate price for it in the docs we could read, so ask your account team how it is billed.

Who hosts what now

Anthropic's docs set out three options:

  • Cloud sandboxes (the default): Anthropic hosts the sandboxes, the service and the model.
  • Self-hosted sandboxes: you host the sandboxes; Anthropic hosts 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.

A deployment has a server (cma-server) that serves the API, runs the agent loop and stores data; a PostgreSQL 16 or later database (without one, the server keeps data in memory and a restart erases it); workers that poll the server and run the agent's shell and file tools in your sandboxes; a console on your server that people sign in to through your identity provider; and an admin CLI that creates environments and their keys. The server has the same API as the hosted service, so the SDK and the ant CLI work against it.

Your choice of model provider

The server's inference.provider setting takes five values: the Claude API (the default), Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry, or a model gateway you run. Agents keep naming models by their Claude API IDs, and the server translates them, so an agent definition does not change when you switch provider. A gateway can accept Claude Messages, Bedrock InvokeModel or Google Cloud rawPredict requests, and can use short-lived tokens that your own callback issues per session.

This is the part most likely to matter for enterprise teams: model calls can stay inside a cloud contract you already have, and with short-lived tokens your callback is told which session, agent and caller each token is for.

What leaves your network

The server never runs the agent's tools itself, but for every model call it sends the conversation so far, including tool inputs and results, to your inference provider. With the default provider, a call to the Claude API also carries the session, agent and environment IDs. Self-hosting the service controls where agents and history are stored; it does not keep the prompt and tool output away from the model provider. If your concern is data residency, pick the provider and region with that in mind.

What the hosted service has that your server does not

The differences page is the document to read before you commit. The gaps that will affect most agent projects:

  • No cloud environments: every environment is self-hosted, so you run workers for all of them.
  • No Files API or file session resources: put files in the worker's working directory, and send images and PDFs as base64. An image or document sent by URL is accepted but the model never sees it, with no error.
  • No webhooks, scheduled deployments or dreams: start sessions from your own scheduler and follow them through the event stream.
  • No prebuilt document skills: upload the docx, pdf, pptx and xlsx skills yourself as custom skills.
  • No mcp_oauth vault credentials: store a static bearer token and rotate it before it expires.
  • Narrower multiagent support: an agent can delegate only to the subagents you list, and dynamic workflows and the advisor are not available.
  • Nothing migrates: agents, skills, vault credentials, memory stores and session history all have to be created again on your server.

Several features are off until you change a setting: MCP servers and Git hosts need allowlist entries (exact host names, no wildcards), and web_search and web_fetch stay off until you enable them. Under some providers the web tools are never offered to the model at all, even though the agent still lists them.

In return, your server has things the hosted service does not: single sign-on through your OpenID Connect provider with roles from group claims, vault secrets in your own secret manager, and short-lived signed agent identity tokens that an MCP server or gateway can check.

What it means for teams building agents

The operational load moves to you. You run and secure the server, its database and the sandboxes, and you install each release yourself. Anthropic's docs are explicit that the agent's bash tool is a real shell with the worker's permissions, and that isolation comes only from the container or VM the worker runs in. Each trust level needs its own deployment.

Plan for silent failures. Several gaps return no error: a tool call waits with no error until a worker polls its environment, a URL image is dropped, a web tool disappears under the wrong provider. In our view, the first job after setup is a test agent that exercises each of these paths and alerts when one goes quiet, alongside the audit trail you would keep anyway.

When not to switch. If your only requirement is that tools run inside your network, Anthropic points you to self-hosted sandboxes, where it still runs the service. If you depend on webhooks, the Files API, dynamic workflows or prebuilt document skills, wait or plan workarounds first. And if you need the model itself on your hardware, this release does not do that: model calls still go to Anthropic or a cloud provider. That is a different decision, covered in our post on running an agent on a model you host.

Frequently asked questions

What is self-hosted Claude Managed Agents?

It is Anthropic's Claude Managed Agents service running on your own infrastructure, in a cloud account or an on-premise data centre. You run the API server, the agent loop, a PostgreSQL database for agents and session history, and the workers that run the agent's tools. Model calls go to the Claude API by default, or to Amazon Bedrock, Claude Models on Google Cloud, Microsoft Foundry or your own model gateway.

How is it different from self-hosted sandboxes?

With self-hosted sandboxes, only the containers or VMs where the agent's tools run are yours, and Anthropic still runs the service and stores the sessions. With self-hosted Managed Agents you host the service too, so agents, sessions and history stay in your database, and you can pick the inference provider. You also take on securing the server and installing each release.

Does any data still go to Anthropic?

With the default provider, yes. For every model call the server sends the conversation so far, including tool inputs and results, to the Claude API, along with the session, agent and environment IDs. If you set the provider to Amazon Bedrock, Google Cloud, Microsoft Foundry or your own gateway, the server needs no Claude API key and model calls go to that provider instead.

How do you get access?

Access is by request rather than self-serve. Anthropic's documentation says to contact your Anthropic account team to request the download, then follow its quickstart to start a server and run a first session. The server software has its own licence agreement, and each release archive includes an operator guide for that version, which Anthropic suggests giving to Claude so it can help with configuration.

Can existing hosted agents move across?

Not automatically. Anthropic's documentation says nothing is copied from the hosted service, including session history, so you recreate agents, skills, vault credentials and memory stores on your server. Check the list of unsupported and changed features first: file resources, webhooks, scheduled deployments, dynamic workflows and prebuilt document skills are not available, and some multiagent and MCP settings need changes.

Share this post

Contact us to build the right product

Talk to our engineers about your application, the systems it connects to, and what you want to build next.

Get in touch
Two people discussing work with a laptop