Self-hosted Claude Managed Agents enters private beta

Anthropic has opened a private beta that lets a company run Claude Managed Agents, its service for running Claude as an autonomous agent, on the company's own cloud account or in its own data center, with model calls going to Anthropic or another provider the company chooses (ClaudeDevs post). The post from ClaudeDevs names the parts that move: the agent harness, the sandboxes, and the session history.

Managed Agents normally runs as a hosted service. Anthropic runs the API and the agent loop, stores agents, sessions and their history, and the agent's tools run in Anthropic's cloud sandboxes, isolated workspaces where the agent executes commands and edits files. Anthropic's engineering write-up explains why it built the service around stable interfaces. An existing beta option, self-hosted sandboxes, already moved tool execution onto customer machines, but Anthropic keeps running the service in that setup and tool inputs and outputs still flow through it. According to Anthropic's documentation, the new option moves the service too: the customer runs a server that serves the same API, runs the agent loop and keeps agents, sessions and history in its own PostgreSQL database, plus workers that run the agent's tools in its own sandboxes. Because the API is the same, Anthropic's SDK and command-line tool work against that server.

The documentation gives two reasons to choose it: an organization must store agents and session history on its own infrastructure, or its model calls must go to Amazon Bedrock, Google Cloud, Microsoft Foundry or its own model gateway. It is also the one form of Managed Agents that the self-hosted sandboxes page lists as eligible for zero data retention. The conversation still leaves the network at each model call, since the server sends the conversation so far, including tool inputs and results, to the chosen model provider. The customer also takes on the work: running and securing the server, database and sandboxes, installing each release, and isolating the agent's shell, which the documentation warns is a real shell with the worker's permissions and network access. Some features exist only on the hosted service, and access goes through an Anthropic account team.