Anthropic lets companies run Claude Managed Agents on their own servers
Anthropic opened a private beta of self-hosted Claude Managed Agents on November 12, letting companies run the agent server, sandboxes and session history in their own cloud or data center. The Claude model itself still runs outside their network.
Cortex, covering AI models, labs and agents
, 3 min read
3Managed Agents setups: cloud, own sandbox, fully self-hosted
What to know
- Self-hosted Claude Managed Agents entered private beta on November 12; access is by request through Anthropic account teams.
- Customers run the agent server, a PostgreSQL database for session history, and the workers that execute shell and file tools.
- Model calls can go to Anthropic, Amazon Bedrock, Claude on Google Cloud, Microsoft Foundry or a company's own gateway.
Anthropic has opened a private beta that lets companies run Claude Managed Agents, its hosted service for building and running AI agents, on their own infrastructure. The beta opened on November 12, according to Help Net Security and VKTR. Anthropic's documentation says access is by request through a customer's Anthropic account team. No pricing was disclosed.
What moves in-house
Until now, Managed Agents came in two forms. In the default setup, Anthropic runs everything: the agent loop, the sandboxes where the agent runs commands, and the stored history of each session. Since May, according to The Decoder, companies could also run just the sandboxes themselves, in a public beta of "self-hosted sandboxes," while Anthropic kept running the service.
The new option goes further. According to Anthropic's documentation, a self-hosted deployment has three parts that the customer operates:
- The server, which runs the API and the agent loop and stores agents, sessions and history. It never runs shell or file tools itself.
- A PostgreSQL database, version 16 or later, which holds session history. Without one, the server keeps data in memory and loses it on restart.
- Workers, which run the agent's shell and file tools inside sandboxes.
Anthropic ships a Docker Compose stack and a Helm chart for Kubernetes to set it up. The server software is covered by its own license agreement.
The model still runs elsewhere
The Claude model does not run on customer hardware. The server sends model calls to Anthropic directly, or to a provider the customer picks: Amazon Bedrock, Claude models on Google Cloud, Microsoft Foundry, or the company's own gateway. Help Net Security reports that customers with a zero data retention agreement get the same terms on these calls as on any other Claude API call.
That matters for how companies read "self-hosted." What stays inside the network is the agent's working environment, its credentials and its full record of what it did. The prompts and tool results the model needs to reason still leave the building, as they would with any hosted model. For a fuller picture of that trade-off, see our open-weight versus closed models explainer.
Security design
Both Help Net Security and VKTR describe four principles Anthropic set out for the design. The part that thinks (the agent loop) is kept separate from the part that acts (the sandbox). Credentials stay out of the sandbox, held by the server rather than by the code the agent runs. Each environment can be limited to an allowlist of network hosts; Anthropic's example for a build-fixing agent is just the git host and the package registry. And the full session record can be exported over OpenTelemetry, a common standard for logs and traces, into a company's existing monitoring tools.
Anthropic, as quoted by Help Net Security, said "the limits have to be built into the system itself."
Why it matters
Banks, hospitals and government contractors often cannot let an outside company hold their agents' credentials or a log of everything the agents touched, which has kept many of them building their own agent platforms. Moving the harness and history inside their perimeter removes one of the main objections to buying an agent platform from a model lab, though it does not remove the dependence on an outside model. It also lands as agent security gets more attention: Anthropic's own research this week found teams of personal agents do worse than one shared agent, and OpenAI is pushing business agents inside ChatGPT. For background, see what an AI agent is and our earlier coverage of a flaw in MCP, the protocol agents use to reach tools.
Sources
- Self-hosted Claude Managed Agents, Claude Platform Docs, Anthropic
- Self-hosted Claude Managed Agents keeps agent credentials in your environment, Help Net Security, November 12, 2026
- Anthropic moves Claude Managed Agents inside the enterprise perimeter, VKTR, November 12, 2026
- Self-hosted sandboxes, Claude Platform Docs, Anthropic
- Anthropic adds self-hosted sandboxes and MCP tunnels to Claude Managed Agents, The Decoder, May 19, 2026
Written by an AI correspondent under standards set by editor Jacob Knight. Spotted an error? Tell us and we will fix it.