Self-Hosted Claude Managed Agents Put the Service On-Site

Anthropic has launched self-hosted Claude Managed Agents, a way for customers to run the whole Managed Agents service on their own infrastructure, not only the sandboxes in which an agent’s tools work. The Claude Platform release notes record the launch on 12 November 2026: “You run the whole Managed Agents service, not only self-hosted sandboxes, on your own infrastructure, and you store agents, sessions, and session history there. The server sends model requests to the provider that you choose, which is Anthropic by default. To request access, contact your Anthropic account team.”

That last sentence is the access route. The self-hosted documentation repeats it, and does not describe a public signup. The customer then runs and secures the stack, and installs each release themselves.

What the customer runs for self-hosted Claude Managed Agents

A deployment, Anthropic says, is one or more servers, their database, and the customer’s sandboxes. The server, cma-server, serves the API, runs the agent loop, and stores agents, sessions and session history. It never runs an agent’s shell or file tools. Those run in workers, either an ant beta:worker command or an SDK worker, inside sandboxes the customer provides. The database is PostgreSQL 16 or later. Without one, the server keeps its data in memory, and a restart erases it.

Model calls leave that stack. By default they go to the Claude API at Anthropic. The inference-provider documentation also allows Amazon Bedrock, Claude Models on Google Cloud (Google managed), Microsoft Foundry, or a model gateway the customer runs. The server holds the inference credential; applications and workers are not given it. For each model call, the server sends the conversation so far, including tool inputs and tool results, to whichever provider is configured.

Anthropic already offered a narrower option, self-hosted sandboxes, in which the customer hosts only the isolated workspaces and Anthropic still runs the service. Self-hosted Claude Managed Agents extends that: the customer hosts the service too. The documentation says this is the path to choose when agents, sessions and history must be stored on the customer’s infrastructure, or when model calls must go to one of those other providers. If the only need is for tools to run on the customer’s machines, the older sandboxes option remains.

The customer’s security duties are correspondingly wider. Anthropic says the agent’s bash tool is a real shell with the worker process’s permissions and network access, and that the worker itself does not isolate that shell. Isolation comes from what the worker runs in. Where the self-hosted-sandboxes security model assigned some jobs to Anthropic, those jobs, including securing the server, hosting memory stores and detecting unusual usage, now sit with the customer. “With self-hosted Managed Agents,” the docs say, “you run and secure a server, its database, and your sandboxes, and you install each release yourself.”

Some features stay on the hosted service

The self-hosted server is meant to offer the same API as Anthropic’s hosted Managed Agents service, so the SDK and the ant CLI can point at it. A differences page lists the exceptions, and Anthropic tells operators to check it before they decide. Some features exist only on the hosted service. Cloud environments are not available; only self_hosted environments exist. The server does not serve the rest of the Claude API, such as the Messages API and the Models API. File session resources and the Files API are rejected. There are no routes for scheduled deployments or “dreams,” and the server sends no webhooks, so a worker that starts on a webhook never starts. Prebuilt document skills for docx, pdf, pptx and xlsx are also rejected; only skills the customer uploads exist. Nothing is copied from the hosted service, session history included, so agents, skills, vault credentials and memory stores have to be created again on the customer’s server.

What the self-hosted option adds, in return, is a choice of inference provider, plus single sign-on, an optional secret manager, and an agent-identity token the server can send with calls to an MCP server.

The launch sits on top of Anthropic’s existing Managed Agents product, which TSN covered when dynamic workflows entered beta. Customers who already run Claude in the enterprise console will recognise the same company security themes TSN has described in Claude Enterprise’s Security Center and inference hooks. Other structured Claude surfaces, such as Smart Reports, remain separate products on Anthropic’s hosted service.

On-site storage is not the same as on-site inference

Self-hosted Claude Managed Agents is a way to keep the service, the database and the sandboxes inside the customer’s network. It is not a way to run the model itself on that network, unless the customer points the server at a gateway they already operate. By default, the conversation still goes to Anthropic. Bedrock, Google Cloud and Microsoft Foundry are alternative destinations, not a local copy of Claude.

The operational burden is also explicit. Access is arranged through the Anthropic account team, not a self-serve button. Once the software is in hand, the customer installs each release, secures the server and the database, and keeps workers running in disposable sandboxes. Some hosted-only features, including webhooks, scheduled deployments and prebuilt office-document skills, will not appear on that server. The licence is Anthropic’s Software License Agreement for the Claude Managed Agents Server, not an open-source grant.

What to watch next is whether organisations that already use self-hosted sandboxes move the service in-house as well, and which inference provider they actually configure. The documentation is written for that choice. It does not, in the pages read for this article, attach a public price or a general availability date beyond the 12 November launch and the instruction to contact the account team.

Sources

• Anthropic, Claude Platform release notes, 12 November 2026 entry on self-hosted Claude Managed Agents.
• Anthropic. “Self-hosted Claude Managed Agents” documentation.
• Anthropic. “Differences from the hosted service.”
• Anthropic. “Inference provider” for self-hosted Claude Managed Agents.

Related stories

• Claude Managed Agents Adds “Dynamic Workflows” in Beta is the nearest TSN account of the hosted Managed Agents product that this self-hosted option extends.
• Claude Enterprise Adds Security Center and Inference Hooks
• Claude Enterprise Gets Smart Reports, With 10 Free Reports a Month