Anthropic's Self-Hosted Claude Managed Agents Enter Private Beta — the Model Still Calls Home
Anthropic has launched self-hosted Claude Managed Agents in private beta, published on 12 November 2026. The agent harness, the sandboxes, the credential handling and the session history can now run inside a customer's own cloud or on-premise — but the model itself does not move. Model requests still go where the customer chooses: to Anthropic directly, to a cloud provider such as Amazon Bedrock, Google's Agent Platform or Microsoft Foundry, or through the customer's own gateway. That distinction is the whole story.
What Claude Managed Agents is
Claude Managed Agents is, in Anthropic's own description, "a suite of composable APIs for building and deploying autonomous agents". Its purpose is to provide the parts of a long-running agent that are hard to build and harder to operate: "the agent harness, sandboxes, credential handling, and session history to get long-running agents into production". A harness is the loop that keeps an agent working through a task. A sandbox is the contained place where the agent runs shell commands and works with files. Credential handling gives the agent the secrets it needs, and session history records what it did, so a long job can be resumed and audited.
Anthropic positions Managed Agents for long-running work in particular. The examples in the announcement are concrete and deliberately unglamorous: fixing broken builds, testing releases, triaging alerts, and working through a bug backlog. None of those is a single chat turn. Each one stretches over time, touches real systems, and needs to keep going when a human steps away. Our earlier coverage of Claude Duet and Claude Commerce Agents explained covers different parts of Anthropic's agent push, while Managed Agents is the underlying set of APIs for building agents of your own. The self-hosted version does not change what Managed Agents is. It changes where it runs.
![]() |
| Modern enterprise server room with rows of rack servers and blinking lights, a technician with a laptop in the aisle (AI-generated illustration, credit: INORYVA) |
What "self-hosted" means here: what runs where
Anthropic's wording is unusually direct: "With self-hosted Managed Agents, every part of the service runs in your own cloud or on-premise, including the harness, sandboxes, credentials, and session history. You operate the Managed Agents server, its database, and the workers where agents run their shell and file tools." That is the definition of "self-hosted" in this announcement. The workers are where agents run their shell and file tools, so they are the part of the system that actually touches the customer's code, files and systems, and keeping them inside the customer's own cloud or on-premise is the point of the offering.
The table below sets out that split in the plainest terms, because the split is the article. Everything on the left runs with the customer. The model call on the right still leaves the customer's infrastructure, by design, and goes to a model provider.
| Part of the system | Where it runs in self-hosted Managed Agents |
|---|---|
| Agent harness, sandboxes, credentials | In the customer's own cloud or on-premise |
| Session history | In the customer's own database |
| Server, database and workers | Operated by the customer |
| Model requests (inference) | To Anthropic, a cloud provider, or the customer's own gateway |
![]() |
| Software developer at a desk with two monitors showing a code editor and a terminal, dark modern office (AI-generated illustration, credit: INORYVA) |
The model still calls home
The part that does not move is the model. Anthropic states: "Model requests go where you choose, whether that's Anthropic directly, your cloud provider (Amazon Bedrock, Google's Agent Platform, or Microsoft Foundry), or your own gateway." The default is Anthropic. None of those options is a model running inside the customer's own infrastructure. Each one is a destination outside the self-hosted service where the model request is answered.
That is why the headline is written the way it is. "Self-hosted Claude Managed Agents" could easily be misread as "Claude, self-hosted" — a Claude model running on a customer's own machines, offline, with no call home. That is not what Anthropic announced. The service is self-hosted. The model is not. OpenAI has its own agent-facing work, covered in our articles on the OpenAI Decisions API and OpenAI offline training, and the industry-wide question is the same: how much of an agent system can live with the customer, and how much still depends on a provider's model. Anthropic's answer here is a clean split.
Session history, access, pricing and the licence
Session history is stored in the customer's own database, and the announcement adds a specific rule for the requests that do leave: "If your organization has zero data retention (ZDR) in place with Anthropic, it applies to the requests sent to Claude in the same way as any other Messages API call." The announcement does not say self-hosting by itself creates a zero data retention arrangement, or that it changes the data terms of model requests. It says the existing arrangement applies in the same way as any other Messages API call. That is a narrower and more precise claim, and it should be read as one.
Access is through the private beta only, via the account team: "Contact your Anthropic account team to request access, or read the docs to learn more." There is no public signup link and no download link in the announcement. There is also no pricing for the self-hosted service and no general availability date, and the announcement names no eligibility criteria, no country list and no named beta customers. The licence states the server binary licence itself is "royalty-free", but "Use of Anthropic Models through any Deployment remains subject to the agreement between the Parties that governs Licensee's access to those models, and nothing in this Agreement changes the pricing, usage policies, or data terms under that agreement". Model calls are still billed under the existing model agreement.
![]() |
| Close-up of hands typing on a laptop keyboard with a glowing agent dashboard on screen (AI-generated illustration, credit: INORYVA) |
What this is, and what it is not
This is a private beta of a self-hosted service, not a general release, and not a step towards a fully offline Claude. The service is the Managed Agents server, its database and its workers, plus the harness, sandboxes, credential handling and session history that run on them. Model requests go to Anthropic by default, or to Amazon Bedrock, Google's Agent Platform or Microsoft Foundry, or through the customer's own gateway. Read that way, the announcement is a careful and genuinely useful step for organisations that want long-running agents to run inside their own infrastructure, with their own credentials and their own session history, while still calling a frontier model hosted elsewhere. The value is in the split itself: more of the agent system under the customer's control, with the model relationship unchanged.
Key Takeaways
- Self-hosted Claude Managed Agents is in private beta, published 12 November 2026, with access by request to the Anthropic account team.
- The harness, sandboxes, credentials and session history run in the customer's own cloud or on-premise; the customer operates the server, its database and the workers.
- Model requests still leave the customer's infrastructure — to Anthropic by default, or to Amazon Bedrock, Google's Agent Platform, Microsoft Foundry, or the customer's own gateway.
- Session history is stored in the customer's own database, and an existing zero data retention arrangement applies to model requests as it would to any other Messages API call.
- The server binary licence is royalty-free, but model use remains billed under the existing model agreement, with pricing, usage policies and data terms unchanged.
- No service price, no general availability date, no public signup link and no named beta customers are in the announcement.
Conclusion
Anthropic's self-hosted Claude Managed Agents private beta is best understood as a change of address for the agent service, not a change of address for the model. The harness, the sandboxes, the credentials and the history can now live with the customer, operated by the customer, in the customer's own cloud or on-premise. The model calls still go out — to Anthropic, or to the cloud provider or gateway the customer chooses — and they are still governed and billed under the existing model agreement. The service is self-hosted. The model still calls home.



Comments
Post a Comment