Self-Hosted Claude Managed Agents: Power-User Guide to Setup, Providers and Limits

Last updated: November 19, 2026

Self-hosted Claude Managed Agents lets you run the whole Managed Agents service on your own infrastructure. Anthropic launched it on November 12, 2026. You host the server, a PostgreSQL database and the sandboxes, store all agents and sessions yourself, and send model calls to Anthropic or another provider. Access comes through your Anthropic account team.

Contents

What is new

The Claude Platform release notes for November 12, 2026 announce that you can now run the whole Managed Agents service, not only the sandboxes, on your own infrastructure. Agents, sessions and session history are stored there, and the server sends model requests to the provider you choose, which is Anthropic by default.

Before this release, the furthest you could go was self-hosted sandboxes, where your containers run the agent’s tools but Anthropic runs the API and storage. The self-hosted overview describes the new option as an extension of that model: you host the service too.

Anthropic does not publish a price for the self-hosted server on these pages. The software has its own license agreement, and the download is available only by request.

Hosting options compared

Anthropic names each option after what you host. The table follows the overview page.

Option Sandboxes Service Model
Cloud sandboxes (default) Anthropic Anthropic Anthropic
Self-hosted sandboxes You Anthropic Anthropic
Self-hosted Managed Agents You You Anthropic, or a provider you choose

Anthropic says to choose the self-hosted service when you must store agents, sessions and history yourself, or when model calls must go to Amazon Bedrock, Claude Models on Google Cloud (Google managed), Microsoft Foundry or your own gateway. If you only need tools to run on your machines, self-hosted sandboxes are simpler.

What you run

One deployment is made of one or more servers, their database and your sandboxes. The overview lists these parts:

  • Server (cma-server): serves the API, runs the agent loop, stores agents and sessions, and calls your inference provider. It never runs shell or file tools.
  • Database: PostgreSQL 16 or later. Without one, data is kept in memory and lost on restart.
  • Workers: an ant beta:worker command or an SDK worker that takes work from the server and runs tools in your sandboxes.
  • Console on your server: a web page where people sign in through your identity provider.
  • Admin CLI (cma-server-admin): creates environments and environment keys on Linux (x86-64 and Arm) and macOS on Apple silicon.

There are three credentials. A server API key authenticates your applications and plays the role a Claude API key plays with the hosted service. An inference credential lets the server call your model provider. An environment key lets workers poll one environment’s work queue.

Security is your job. The agent’s bash tool is a real shell with the worker’s permissions and network access, and the worker does not isolate it. Isolation comes from the container or VM the worker runs in.

How to try it with Docker Compose

The quickstart runs the stack on one machine. You need the release archive, a Claude API key, current Docker Engine with the Compose plugin, at least two CPUs, and openssl plus the ant CLI.

  1. Extract the archive and change into its compose directory.
  2. Create a .env file with a database password, your Claude API key, a generated server API key and a credentials encryption key, then chmod 600 .env.
  3. Run docker compose up --build -d and wait for the log line cma-server listening.
  4. Check curl -s http://127.0.0.1:8400/cma-server/readyz until it reports "status":"ready".
  5. In a second terminal, set ANTHROPIC_BASE_URL to http://127.0.0.1:8400 and ANTHROPIC_API_KEY to the server API key, not your Claude key.
  6. Create an agent with ant apply --yes quickstart-agent.md, start a session in the compose-sandbox environment, and connect with ant beta:sessions connect.
  7. Ask the agent to run hostname. A reply starting with cma-sandbox- confirms the tool ran in a sandbox container.

Anthropic says the Compose stack is for development, not production. It serves plain HTTP on port 8400. For production it recommends the two Helm charts in the release, cma-server and cma-sandboxes; the Kubernetes guide asks for Kubernetes 1.30 or later, gVisor nodes and the Agent Sandbox controller.

Choosing an inference provider

According to the inference provider page, the inference.provider setting controls where model calls go:

inference.provider Calls go to Authentication
anthropic (default) Claude API ANTHROPIC_API_KEY
bedrock Amazon Bedrock Your AWS credential
vertex Claude Models on Google Cloud (Google managed) Google Application Default Credentials
foundry Microsoft Foundry Foundry API key or Microsoft Entra ID
gateway A model gateway you run A key or token your gateway accepts

Agents keep using Claude API model IDs such as claude-opus-5-5, and the server translates them for the provider. On Bedrock, the server adds the us.anthropic. cross-region prefix by default and needs invoke and list permissions on the models it uses.

For each model call, the server sends the conversation so far, including tool inputs and results, to your provider.

Differences from the hosted service

The differences page says the API is the same apart from listed exceptions, and nothing is copied over from the hosted service. You recreate agents, skills, vault credentials and memory stores.

  • Only on your server: a choice of inference provider, single sign-on through OpenID Connect, your own secret manager, and signed agent identity tokens for MCP calls.
  • Not available: cloud environments, the Messages and Models APIs, the Files API, mcp_oauth vault credentials, prebuilt document skills, scheduled deployments, dreams and webhooks.
  • Different: multiagent blocks must use type coordinator with 1 to 20 listed agents; images sent by URL are silently ignored; MCP results are cut at 100 KiB by default.
  • Off until configured: MCP hosts and Git hosts need allowlist entries, and web_search and web_fetch must be turned on in settings.

Workflows, limitations and alternatives

Anthropic says each trust level needs its own deployment. Replace scheduled deployments with your own cron job, and keep workers running because there are no webhooks. Base64-encode images and PDFs before sending them.

The main costs are operational: you install each release, secure the server and database, host memory stores and watch for unusual usage.

For related Claude Platform changes, see our guides to Claude Fable 5.2, Claude Linked billing and the Claude IP allowlist. For other managed agent platforms, compare Mistral Managed Agents and the OpenAI Agents API.

Frequently asked questions

What is self-hosted Claude Managed Agents?

It is Anthropic’s Managed Agents service running on your own infrastructure, in a cloud account or an on-premise data center. You host the server, its database and the sandboxes. Agents, sessions and session history are stored on your side, and the model runs at Anthropic or at a provider you choose.

How do I get access to self-hosted Claude Managed Agents?

Contact your Anthropic account team. The release archive is not a public download. If your Claude Console organization already has access, sign in to the Console, switch to that organization and reload the docs page.

How is it different from self-hosted sandboxes?

With self-hosted sandboxes, only the tool execution environment runs on your side and Anthropic runs the service. With self-hosted Managed Agents, you also run the API, the agent loop and the storage for agents and sessions.

Can self-hosted Managed Agents use Amazon Bedrock or Google Cloud?

Yes. The inference.provider setting accepts anthropic (the default), bedrock, vertex for Claude Models on Google Cloud (Google managed), foundry for Microsoft Foundry, and gateway for a model gateway you run.

Which hosted features are missing on a self-hosted server?

Anthropic lists cloud environments, the Messages and Models APIs, the Files API, mcp_oauth vault credentials, the prebuilt docx, pdf, pptx and xlsx skills, scheduled deployments, dreams and webhooks as unavailable on your own server.

Sources

Image: Claude logo by Anthropic, public domain (trademarked), via Wikimedia Commons. Claude is a trademark of Anthropic.

Related answers

Leave a Comment