Sunday, November 15, 2026 Independent reporting on artificial intelligence Newsletter
AI News Beats
AI for Business

Anthropic opens private beta for self-hosted Claude agents

Businesses in the beta run the agent service and store session history themselves, while choosing where requests to the language model are sent.

Two tall racks filled with electronic components and bundled wires
Photo: Eric Stoynov / Unsplash

Anthropic opened a private beta of self-hosted Claude Managed Agents on November 12, giving selected organizations a way to run the agent service in their own cloud or on their premises. The change is aimed at teams that want more control over software that interacts with internal systems and may keep working after a person has signed off.

The hosted version puts Anthropic in charge of the underlying service. The company had already let customers host their own sandboxes, the isolated environments where agent tools run. In the new setup, Anthropic says the customer also operates the server that coordinates the agent, the database that keeps its session history and the workers that run the tools.

Self-hosting does not mean local model inference

Customers can choose where requests to Claude go. Anthropic lists its own service, Amazon Bedrock, Google’s Agent Platform, Microsoft Foundry or a gateway the customer operates. The location of the agent service and the destination of its model requests are therefore separate decisions. An organization can hold session records in its own database while still sending some request data to an external model provider.

That distinction matters when evaluating retention and confidentiality. Anthropic says an existing zero data retention arrangement applies to Claude requests sent to Anthropic in the same way it does to other Messages API calls. An organization using another route still needs to establish which information it sends and which provider terms apply. The announcement does not say that hosting the agent eliminates the need for that review.

More control also means more responsibility

The company describes an architecture that separates the agent’s coordination service from the environment where its tools execute. It says credentials are kept outside that execution environment and that customers can use their own secrets manager. Anthropic also says customers can restrict network destinations and export session activity to their monitoring tools through OpenTelemetry.

Operating the service internally shifts day-to-day work to the customer. A team has to decide who will handle database availability, review agent activity and respond when a worker or model connection fails. For an unattended job, a useful trial would compare the new deployment’s reliability with the hosted version and check whether operators have the information they need to investigate a bad result.

Availability

Anthropic says developers use the same core APIs as on its hosted Managed Agents service. Keeping the interface familiar could make a trial easier for an existing customer, but switching where the service runs still changes who maintains its infrastructure and controls. The work needed to migrate would depend on the systems an agent already uses.

The company is directing interested customers to their Anthropic account teams for beta access. It did not announce a public launch date or standalone beta pricing. Broader availability and the ongoing cost of running the service will determine how widely businesses can adopt this option.

Stay on top of AI

Get the stories that matter in artificial intelligence, straight to your inbox.

Free. One email with the stories that matter. Unsubscribe any time. See our privacy policy.