Dev & code

Anthropic allows self-hosting Claude Managed Agents: the entire agent service runs on-premise, except the model

Since November 12, in private beta, enterprises can run the entire Claude Managed Agents service (agent loop, sandboxes, secrets, session history) within their own cloud or on-premise servers. The model remains hosted by Anthropic or a chosen provider (Bedrock, Google Cloud, Foundry, custom gateway): each call sends the full conversation to it. No pricing has been published.

A glass fortress housing internal gears, connected by a glowing cable to a distant luminous orb.
AI-generated illustration

An agent that fixes builds, triages alerts, or works through a bug backlog needs access to an organization’s code and internal systems, often without supervision. For many security teams, entrusting that execution loop to a third-party hosted service remains a major obstacle. Anthropic responded on November 12 with a self-hosted version of Claude Managed Agents, its suite of APIs for building and running autonomous agents. Currently in private beta, prospective users must go through Anthropic’s sales team to obtain the software.

What changes: the entire service moves to the customer

Until now, Anthropic offered two configurations: fully hosted on its cloud, or keeping only the sandboxes (the containers where the agent runs commands) on the customer side. This third option moves the “brain” as well: the API, the agent loop, secrets, and session histories. According to the documentation, a deployment includes a cma-server, a PostgreSQL 16 or newer database, workers executing shell and file tools, a web console connected to the enterprise identity provider, and an administrative CLI. The API remains identical to the hosted service: the SDK and the ant CLI plug into it without modification.

Anthropic highlights four core design principles across all three options: the agent loop is separated from the sandbox where code executes; credentials stay out of the sandbox, injected via a proxy; network access can be restricted to an allowlist of permitted hosts; and every session leaves a complete audit log, exportable via OpenTelemetry to existing monitoring tools.

What stays the same: the model remains outside

The crucial detail not to overlook: Claude does not run locally. The server routes its requests to the Claude API by default, or to Amazon Bedrock, Claude on Google Cloud (managed by Google), Microsoft Foundry, or an in-house gateway. Crucially, with every call, it transmits the entire conversation, including tool inputs and outputs. Self-hosting therefore governs where histories and secrets are stored, not what the model provider sees. A Zero Data Retention (ZDR) agreement with Anthropic applies to these requests just like standard API calls.

Customers also take on operational responsibilities. The documentation is explicit: the agent’s bash tool is a real shell carrying the worker’s permissions and network access, and the worker does not provide isolation. That responsibility falls entirely on the container or virtual machine, and separate trust levels require separate deployments. Customers must also install each update themselves. The software is released under a proprietary license, and no pricing has been disclosed. According to The Next Web, certain hosted features, such as webhooks and scheduled triggers, are not yet supported.

Our take

Anthropic is joining an established trend: in early November, Mistral announced that its managed agents would arrive in private cloud and on-premise environments by late November. The difference is significant: Mistral promises that the model itself can run on-premise, whereas Anthropic keeps Claude outside the customer’s network perimeter. As a result, this offering addresses a specific need—maintaining control over secrets, audit logs, and agent history while choosing one’s cloud provider—without meeting the requirement for fully internal data processing. For teams considering adoption, the key question is not “where does the agent run?” but “what data is being sent to the model?”, accompanied by the burden of securing and patching a server that orchestrates agents equipped with shell access.

Sources

  1. Claude — Operate Claude Managed Agents inside your security perimeter (announcement) Primary source
  2. Claude Platform Docs — Self-hosted Claude Managed Agents Primary source
  3. Claude Platform Docs — Release notes (November 12, 2026) Primary source
  4. The Next Web — Anthropic lets companies run Claude's agent service on their own servers (November 12, 2026) Secondary source

A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.