Self-hosted Claude Managed Agents: what you run, and what stops working
Anthropic opened a private beta of self-hosted Claude Managed Agents on November 12: the server, database and sandboxes run in your own cloud or data center. Its docs also list the hosted features you lose.
Anthropic opened a private beta of self-hosted Claude Managed Agents on November 12. Until now, teams could move only the sandbox, where an agent's tools run, onto their own machines. With this option the whole service runs in your own cloud account or data center: the API, the agent loop, and the storage for agents, sessions and session history. Model calls can go to Anthropic, to Amazon Bedrock, Google's Agent Platform or Microsoft Foundry, or to your own gateway. Access is by request through an Anthropic account team, and the server software comes under its own license agreement.
The docs describe what you take on. A deployment is a cma-server process that serves the API and runs the agent loop, a PostgreSQL 16 or later database (without one, the server keeps data in memory and a restart erases it), and workers that run the agent's shell and file tools. A worker is an ant beta:worker command or an SDK worker. There is also a console page on your server with sign-in through your identity provider, and an admin CLI, cma-server-admin, for Linux and Apple silicon Macs. The docs warn that the worker does not isolate the agent's bash tool; isolation comes from the container or VM the worker runs in. You also install each release yourself.
What changes if you move an existing app
The server speaks the same Managed Agents API, so the SDK and the ant CLI point at it, but Anthropic's differences page lists gaps worth checking before you move. Nothing is copied over from the hosted service, including session history.
- Only self_hosted environments exist; asking for a cloud environment returns a 400 error.
- The server does not serve the rest of the Claude API: SDK calls to /v1/messages and /v1/models return 404, so send those from a separate client.
- No Files API or file session resources, no mcp_oauth vault credentials, and no prebuilt docx, pdf, pptx or xlsx skills; you can upload those skills from the anthropics/skills repository as custom skills.
- No scheduled deployments, dreams or webhooks; start sessions from your own scheduler and follow them through the event stream.
- MCP servers and Git hosts must be added to allowlists that start empty, and the web_search and web_fetch tools are off until you enable them.
- An image or document sent by URL is accepted but the model never sees it, with no error; send it as base64 instead.
In return, only the self-hosted server offers a choice of inference provider, single sign-on through an OpenID Connect provider, storing vault secrets in your own secret manager, and short-lived signed tokens that identify an agent to your MCP servers. If you only need tools to run on your own infrastructure, the docs point to self-hosted sandboxes instead, where Anthropic still runs the service.
Try it with your AI. If you already build on hosted Managed Agents, this prompt has your assistant audit your code against the self-hosted gaps before you request access:
I build on Claude Managed Agents and am considering the self-hosted server. Audit my
code for features that behave differently or are missing there. Do not change any files;
report first.
1. First, find where my code creates agents, environments and sessions: [path to my
agent code].
2. List every place that uses any of these, with file and line: cloud environments;
calls to /v1/messages or /v1/models through the same client; the Files API or file
session resources; mcp_oauth vault credentials; prebuilt docx, pdf, pptx or xlsx
skills; scheduled deployments; dreams; webhooks; image or document blocks with a url
source; MCP servers; github_repository resources; web_search or web_fetch.
3. For each, say what the self-hosted docs suggest instead, quoting
https://platform.claude.com/docs/en/managed-agents/self-hosted/differences. If you
cannot read that page, say so and stop rather than guessing.
4. List the MCP and Git hostnames I would need to allowlist.
Show me the report and wait for my go-ahead before proposing any code changes.Source: Anthropic: Operate Claude Managed Agents inside your security perimeter, Claude docs: Self-hosted Managed Agents, Claude docs: Differences from the hosted service
This story first appeared in the claudefox daily briefing for November 14, 2026.