Running Claude agents in-house does not make the model local
Anthropic’s November 12 platform release notes announced self-hosted Claude Managed Agents: customers can operate the entire service on their own infrastructure, not just the environment where tools execute.
Help Net Security described the option as a private beta, with access requested through an Anthropic account team.
This is a development announced on November 12, not a product launched this Sunday.
The technical distinction involves three layers.
Sandboxes are the isolated workspaces where tools run commands and handle files; the service includes the API, agent loop and storage for agents, sessions and session history.
The self-hosted option moves both layers onto customer infrastructure, whether a cloud account or an on-premise data centre.
The model still runs at Anthropic or at a provider selected by the customer.
Keeping session history in-house does not mean every piece of information stays inside the network.
The documentation says each model call sends the conversation so far, including tool inputs and results, to the inference provider; the default destination is the Claude API.
The deployment changes where the service and its records live, but does not automatically create an offline AI system or a locally installed model.
Greater control also brings operational responsibility.
Customers must run and secure the server, database and sandboxes, and install releases themselves.
The guide warns that the bash tool is a real shell with the worker process’s permissions and network access: the worker does not isolate it by itself, so protection depends on the environment around it.
Moving the software is therefore not proof of security; isolation, permissions and maintenance remain part of the deployment.
Access to the download requires contacting Anthropic’s account team, and the server software is covered by its own contractual licence.
This should not be described as an open-source release or immediate universal availability.
The sources establish the architecture and access route, but this article includes no performance testing, compliance assessment or cost comparison, and offers no certification of security or value for money.
