Operate Claude Managed Agents inside your security perimeter

Open original

Anthropic now lets companies run the full Claude Managed Agents service inside their own cloud or on-premise environment, including the agent harness, credentials, and session history. Model requests can still go to Anthropic, a cloud provider like Bedrock or Vertex, or a custom gateway.

  • Previously only the sandbox could be self-hosted; now the whole control plane, including the server, database, and workers, runs in your environment.
  • Session history lives in your own database, and zero data retention agreements apply to Claude requests the same way as standard Messages API calls.
  • Security is built into the architecture: the agent's code runs in a sandbox separate from the harness and secrets, network access is limited to an allowlist of hosts, and every tool call is logged and exportable via OpenTelemetry.

Anthropic has put out a private beta of self-hosted Claude Managed Agents, which lets a company run the entire agent service inside its own cloud or on-premise environment rather than on Anthropic's hosted platform. Until now, teams that wanted their code kept in-house could only self-host the sandbox, where the agent's shell and file tools run. Now the harness, credentials, session history, and the Managed Agents server and database can all live on the customer's side too. Companies choose where model requests go: directly to Anthropic, through Amazon Bedrock, Google's Agent Platform, or Microsoft Foundry, or through their own gateway. Developers use the same core APIs as the hosted version, so the agents they build should move between the two without rewriting. Session history sits in the customer's own database, and if an organization has zero data retention in place with Anthropic, that applies to Claude requests the same way it does for any Messages API call. The reason for the push is that agents are doing more consequential work now, such as fixing broken builds, testing releases, triaging alerts, and chewing through bug backlogs, usually unattended and with access to internal systems. The article's argument is that when nobody is watching each step, security has to be designed into the architecture. Managed Agents is built around four ideas. The "brain" (the harness that calls Claude) is separated from the "hands" (the sandbox where code runs), so a sandbox crash doesn't kill a run, and sandboxes only spin up when a tool needs one. Credentials are kept out of the sandbox entirely: for MCP tools and supported API calls, a proxy outside the sandbox attaches the secret to the request, and self-hosters can use their own secrets manager. Each environment can be restricted to an allowlist of hosts, with everything else blocked; the example given is a build-fixing agent that might only reach a git host and a package registry. Finally, every step, including tool calls and their results, is logged and exportable over OpenTelemetry into existing monitoring tools. What the article doesn't tell you is how mature the private beta is, what the operational burden of running the server, database, and workers actually looks like, or how pricing works for self-hosted deployments. Access requires contacting an Anthropic account team, and the docs are the place to find the implementation details that the marketing page leaves out. For a team already running agents against internal code and wanting the control plane under its own roof, this is the piece that was missing. For everyone else, the more useful takeaway is the architecture itself: the separation of brain and hands, credentials kept out of reach, and default-deny networking are sensible patterns regardless of which vendor you use.

Ask about this article