Anthropic opened a private beta on November 12, 2026 that moves its agent platform inside the customer’s own perimeter. Self-hosted Claude Managed Agents lets an organisation run the agent harness, the sandboxes and the session history on its own infrastructure, in its own cloud account or on-premise, while model calls go to whichever provider it chooses.
What actually moves
Managed Agents is Anthropic’s suite of APIs for building and running autonomous agents: you define an agent, start a session, and exchange events while it works. Until now the service ran on Anthropic’s infrastructure, with an earlier option to self-host only the sandboxes where an agent’s tools execute.
The new option moves the service too. According to Anthropic’s documentation, a self-hosted deployment includes the API, the agent loop and the storage for agents, sessions and session history, plus the server, its database, the workers that run shell and file tools, and the sandboxes. The company’s comparison table is blunt about the trade: with cloud sandboxes, Anthropic hosts everything; with self-hosted sandboxes, you host execution but Anthropic still runs the service; with self-hosted Managed Agents, you host the service and choose the model provider.
Why teams asked for it
Anthropic’s own framing is that agents have moved from demos to work that touches code and internal systems: fixing broken builds, testing releases, triaging alerts, clearing a bug backlog. That work needs credentials and access, and it runs unattended. For regulated industries, or any company with data-residency rules, the sticking point was never the model call alone. It was where the credentials, the session logs and the agent’s memory lived.
Self-hosting answers that with a straightforward trade: you keep the data and you take on the operational load.
Where the model calls go
This is the part worth reading twice. Self-hosting the service does not mean self-hosting the model. Requests can go to Anthropic directly, to Amazon Bedrock, to Claude on Google Cloud, to Microsoft Foundry, or to a gateway the customer operates. The company says that if an organisation has zero data retention in place with Anthropic, it applies to those requests in the same way as any other Messages API call.
Anthropic also says every session leaves a record of the steps an agent took, including tool calls and their results, and that the record exports over OpenTelemetry so it can land in existing monitoring tools. For anyone who has tried to audit an agent after the fact, that is the feature that matters most.
What it does not solve
Three limits are explicit in the documentation. You run and secure the server, its database and its sandboxes, and you install each release yourself, so patching becomes your job. Some features exist only on the hosted service, and Anthropic tells readers to check the differences page before committing. And access is gated: this is a private beta for eligible organisations, requested through an Anthropic account team, not a self-serve download.
There is also a conceptual limit. Self-hosting changes where the plumbing runs, not what the model can do. Teams weighing the option should compare it against self-hosted sandboxes first; if the only requirement is that tool execution stays inside the network, the lighter option may be enough.
How it fits with the rest of the platform
The same APIs and SDKs work against a self-hosted server, which keeps migration costs down and means existing agent definitions are not stranded. The release notes list the beta alongside other November changes, including memory stores attached to self-hosted sandboxes, so agents can read and write persistent context inside the sandbox.
For readers tracking Anthropic’s model releases rather than its infrastructure, our breakdown of Claude Fable 5.2 pricing and who should use it covers the model side of the same platform.

Sources
- Anthropic: Operate Claude Managed Agents inside your security perimeter
- Claude Platform docs: Self-hosted Claude Managed Agents
- Claude Platform release notes, November 12, 2026
- Claude Devs: self-hosted Managed Agents announcement





