Anthropic lets companies run "Claude Managed Agents" in their own cloud or on premises; private beta begins

Companies can run the agent platform themselves

Anthropic said on Nov. 12 (Nov. 13 Japan time) that it has begun offering "Self-hosted Claude Managed Agents" in private beta, letting companies run Claude Managed Agents, its platform for building AI agents, in their own cloud or on-premises environments. ClaudeDevs, the official X account for people building with Claude, posted the announcement at 2:57:14 a.m. Nov. 13 Japan time.

According to the post, companies can run three things themselves: the core system that runs the agent (the harness), the isolated environments the agent works in (sandboxes) and session history. They can also choose where requests to Claude are sent.

The official documentation describes Managed Agents in two parts. One is the sandboxes, where an agent's tools run commands and work on files; the other is "the service," which covers the API, the system that drives the agent's work (the agent loop) and storage for agents, sessions and their history. With Self-hosted Claude Managed Agents, the company runs both, and the model runs at Anthropic or at a provider the company chooses.

Companies can choose where requests to Claude go

According to the documentation, model requests can be sent to Amazon Bedrock, Claude models on Google Cloud (Google managed), Microsoft Foundry or a model gateway the company runs itself. The default destination is the Claude API.

Each time the company's server calls the model, it sends the conversation so far, including tool inputs and results, to that destination. Calls to the Claude API also carry the IDs of the session, agent and environment.

Companies run and secure the server, the database and the sandboxes themselves, and install each new release on their own. The database is PostgreSQL 16 or later; without a database, data is kept in memory and erased on restart.

The scope widens from May's "self-hosted sandboxes"

On May 19, Anthropic began offering "self-hosted sandboxes" in public beta, which moved only tool execution in Managed Agents to the customer's environment. At that time, the agent loop stayed on Anthropic's infrastructure.

The new option moves the service to the customer's side as well. The documentation directs companies that only need tools to run on their own infrastructure to self-hosted sandboxes, where Anthropic runs the service.

Google said on Nov. 11 that Gemini API's identically named feature, "Managed Agents," can now connect to customers' own sandboxes. Google's feature is a separate product from Anthropic's.

Access goes through account teams, and some features are unavailable

To use it, companies contact Anthropic's account team to request access to the software download. Pricing and eligibility requirements are not given in the post on X or in the body of the documentation.

A company's own server has the same API as the service Anthropic runs, but some features are unavailable. Cloud sandboxes provided by Anthropic, the Files API for handling files, prebuilt skills for Word, PDF, PowerPoint and Excel documents, scheduled runs and webhook notifications cannot be used. Nothing carries over from Anthropic's service, session history included, and agents and other resources must be created again on the company's server.