Anthropic now lets companies run the whole Claude Managed Agents service on their own infrastructure, not just the sandboxes where agents run code. It is a private beta announced on 12 November, access goes through an Anthropic account team, and there is no public price. Read the license first: one clause gives you 8 hours to roll back if Anthropic flags a safety risk.
What you host and what still leaves
Managed Agents is Anthropic’s set of APIs for long-running agents: the agent loop that calls the model, sandboxes, credential handling and session history. Before this, you could host only the sandboxes. Under the self-hosted option, you also run the server, its database and the workers that execute shell and file tools.
Model calls still leave your network, to the Claude API by default or to Amazon Bedrock, Google Cloud, Microsoft Foundry or your own gateway. Each call carries the conversation so far, tool results included. The data and access page says calls to the Claude API also carry session, agent and environment IDs, and no setting turns them off.
For teams whose agents touch private repos and build systems, the gain is real. Session history sits in your database and credentials can live in your own secret manager. Our plain explainer on MCP connections explains how agents reach those internal tools in the first place.
Setup time and what you give up
The quickstart runs on one machine with Docker Compose: the server, PostgreSQL and 10 workers, each serving one session at a time. Anthropic calls that stack development only and points production users to Helm charts for Kubernetes. Without a database, the server keeps its data in memory and a restart erases it.
Moving over is a rebuild. The differences page says nothing is copied from the hosted service, so agents, skills, vault credentials and memory stores are created again. Missing on your server:
- webhooks and scheduled runs, so sessions start from your own scheduler, such as cron
- the Files API, so files go in the worker’s working directory
- OAuth credentials for MCP servers, so you store a static token and rotate it by hand
- the prebuilt Word, PDF, PowerPoint and Excel skills, unless you upload them as custom skills
In return you get single sign-on through your own identity provider.
The license is where it gets tight
The server license ships a binary you may not modify or reverse engineer, and bars non-Anthropic models unless the docs allow them. After a written safety, security or misuse notice, you must disable the feature or roll back in every deployment within 8 hours. Designated safety updates go in within 1 business day. Regulators are watching agent safeguards too: California is suing OpenAI over agents that escaped testing and attacked Hugging Face.
Anthropic may also reserve some Claude models for its hosted service, and you may not block the usage signals the software sends back. Patrick Moorhead, chief analyst at Moor Insights & Strategy, told The New Stack: “Self-hosted usually means the customer owns the change window. This license flips that.”
What it costs
The pricing page bills the hosted service for tokens plus $0.08 per session-hour of running time. For self-hosting it says to contact your account team. Your own bill adds servers, PostgreSQL 16 or later, and someone on call who can act on an 8-hour notice.
That on-call duty is the hard part for a small team already running agents for chores like AWS Kiro’s recurring coding bots. We have not run the server; the download is limited to beta customers, so everything here comes from Anthropic’s docs and license.
What to do this week
If security review is what keeps agents off internal systems, send your security team the data page and the license before asking for access. Check whether you can roll back a production service within 8 hours and which missing features your agents use. If you only need tools to run in-house, self-hosted sandboxes keep Anthropic running the server.
Sources
Get the weekly email
The week's most useful AI news, a tutorial and an interview.




