Anthropic launched self-hosted Claude Managed Agents on November 12, 2026, with access offered through a private beta. Customers can run the agent service, sandboxes and session storage in their own cloud or on-premises infrastructure. Model inference remains at Anthropic or the selected provider, a boundary that matters for universities assessing where research material is processed.
Self-hosting expands beyond the sandbox
The November 12 release notes confirm that the launch covers the whole Managed Agents service. Previously, self-hosted sandboxes let customers run agent tools on their infrastructure while Anthropic operated the service. The new option also moves the API, agent loop and storage for agents, sessions and their history to the customer.
In its launch announcement, Anthropic says customers operate the server, database and workers that execute shell and file tools. Developers retain the same core APIs. Access requires contacting an Anthropic account team; the announcement describes private beta availability.
Local session storage still involves external inference
Anthropic’s data and access documentation states that each model call sends the conversation so far, including tool inputs and results, to the inference provider. The default is the Claude API. The deployment overview lists other choices including Amazon Bedrock, Claude Models on Google Cloud (Google managed), Microsoft Foundry and a customer’s model gateway. The organization’s agreement with the inference provider governs retention of model inputs and outputs.
For university research data governance, this creates separate questions about custody of stored records and disclosure during processing. Keeping session history in an institutional database does not mean all research content stays on premises. The launch supplies a deployment option for the agent service, without supplying locally hosted model weights or an air-gapped Claude service.
Universities inherit operating responsibilities
The deployment overview assigns customers responsibility for securing the server, database and sandboxes, installing releases, hosting memory stores and detecting unusual usage. It also says the worker’s shell has the worker process’s permissions and network access: isolation comes from the container or virtual machine around it.
For research offices evaluating the beta, the resulting division of responsibility is:
- Record custody: the institution operates session storage, bringing agent records into its own infrastructure-management arrangements.
- Model processing: the selected provider receives conversation content needed for inference, so local storage alone does not settle the data-transfer question.
- Tool execution: the institution operates the workers and their isolation, making permissions and network boundaries part of the deployment decision.
This is an institutional interpretation of the documented architecture. It connects the launch to the accountability questions discussed in CASRAI’s coverage of agentic AI governance.
The beta has migration and feature limits
Anthropic’s differences page says hosted resources, including session history, are not copied into a self-hosted deployment. Agents, skills, vault credentials and memory stores must be created again. Hosted cloud environments, the Files API, scheduled deployments and webhooks are unavailable. Web search and web fetch require server configuration, while MCP connections require an allowed-host entry.
These limits make application compatibility a separate consideration from control over storage. The development concerns where the agent platform runs, distinct from the November 9 Claude model release. For researchers and university IT teams, the immediate change is a private-beta route to operating that platform themselves, together with the duties that operation entails.
Sources
- Anthropic: Claude Platform release notes, November 12, 2026. Checked 2026-11-14.
- Anthropic: Operate Claude Managed Agents inside your security perimeter, November 12, 2026. Checked 2026-11-14.
- Anthropic: Self-hosted Claude Managed Agents. Checked 2026-11-14.
- Anthropic: Data and access in self-hosted Claude Managed Agents. Checked 2026-11-14.
- Anthropic: Differences from the hosted service. Checked 2026-11-14.
