Your AI agent just phoned home. That is the hidden cost of every cloud-based automation you have been running. The agent does the work, yes. But the credentials it uses, the session history it logs, and the data it touches all pass through someone else’s infrastructure. For most businesses, that tradeoff has been invisible. Anthropic just made it optional.
On November 12th, Anthropic opened a private beta for self-hosted Claude Managed Agents. The idea is straightforward: instead of running your agents on Anthropic’s servers, you run the entire orchestration stack inside your own cloud or on-premises environment. The sandbox, the credential handling, the session database, the logs. All of it stays within your security perimeter.
What “Self-Hosted” Actually Means Here
Think of it as the difference between hiring a contractor who works in your office versus one who takes your files home. The work gets done either way. But only one setup lets you watch exactly what happens to your data.
The technical design separates the control layer from the execution layer. The Claude model itself still calls back to Anthropic (or a supported provider you choose), but every other piece of the system runs in your environment. Your credentials stay in your own secrets manager. Network traffic is locked to an allowlist of permitted hosts you define. Every tool call and its result gets logged and can feed directly into your existing monitoring systems via OpenTelemetry, which is the standard protocol most engineering teams already use to track what their applications are doing.
This is a meaningful upgrade from what Anthropic offered in May 2026, when only the sandbox execution environment could be self-hosted while the agent loop itself still ran on Anthropic’s infrastructure. The November private beta moves the whole Managed Agents service into the customer environment.
Who Needs This Most
Two types of organizations stand to gain the most. The first is any company handling regulated data: healthcare records, financial information, or anything covered by HIPAA, SOC 2, or GDPR. The second is any company that has had the conversation with their legal or IT team and landed on “our data stays on our servers” as a firm requirement. Until now, those organizations were largely locked out of the most capable AI agent workflows. That changes with this release.
Access is currently by request through an Anthropic account team. It is not open to everyone yet. But the direction is clear. The enterprise AI market is moving toward deployment models where customers own and control their infrastructure, and the most capable AI labs are starting to meet that demand.
For businesses that have been watching AI automation from the sidelines specifically because of data privacy concerns, this is a development worth taking seriously. The capability gap between what you could do with cloud AI and what you could run yourself has been real for years. It is closing fast.
Want to explore how self-hosted AI agents could benefit your business? Let’s talk.