CoAI
News

Anthropic Releases a Version of Its AI Agent Platform "Claude Managed Agents" That Runs in Your Own Environment, Keeping Conversation Records In-House

On November 12, 2026 (US time), Anthropic released a self-hosted version of "Claude Managed Agents," its service for building and running AI agents, that companies can run in their own cloud or data center. Agent settings and conversation records are stored in the company's own database, and models can come from Anthropic or from providers such as Amazon Bedrock. It is first offered as a limited beta.

November 12, 20263 min read
Screenshot of the official Anthropic — Operate Claude Managed Agents inside your security perimeter page
Anthropic — Operate Claude Managed Agents inside your security perimeter · SourceCaptured: November 14, 2026

Screenshots show the official pages on the capture date and may differ from the pages at publication. Select an image to enlarge it.

On November 12, 2026 (US time), Anthropic released a "self-hosted version" of "Claude Managed Agents," its service for building and running AI agents, that companies can run in their own environments. The entire service can be placed in a company's own cloud account or its own data center. It is first offered as a limited beta, and using it requires applying through Anthropic's sales team.

Running the entire service in-house

Claude Managed Agents is a service for building and running AI agents powered by Claude through a structured API. It originally ran on Anthropic's servers, and later companies became able to host only the place where agents run programs and work with files (the sandbox) themselves.

With the new self-hosted version, companies run not only the sandbox but also the system that runs the agents. The differences are as follows.

  • Standard version: Anthropic runs the sandbox, the service and the model
  • Self-hosted sandbox: only the sandbox is in-house. The service and the model are Anthropic's
  • New self-hosted version: the company runs the sandbox and the service. The model comes from Anthropic or a provider the company chooses

The API is the same as the standard version, so existing development kits and command-line tools work as they are.

Data that stays in-house and data that is sent out

Agent settings, sessions (the unit of work) and their records are stored in the company's own database (PostgreSQL 16 or later) and remain until the company deletes them. Anthropic says it cannot connect from outside to the company's servers or their records.

On the other hand, each time an AI model is called, the conversation so far and tool inputs and outputs are sent to the model provider. The provider can be chosen from the following.

  • Anthropic's Claude API (default)
  • Amazon Bedrock
  • Google Cloud
  • Microsoft Foundry
  • A model relay server the company runs itself

However, some features, such as Web search and reading Web pages, may not be available depending on the provider chosen.

Required environment and limitations

For production environments, running it on Kubernetes (1.30 or later) is recommended. If you just want to try it, you can also bring it up on a single machine with Docker Compose. Securing the servers, applying updates and detecting suspicious use are the responsibility of the company using it.

The following features of the standard version are not available in the self-hosted version.

  • Other Claude API features, such as Files
  • Ready-made skills for working with Word, PDF, PowerPoint and Excel
  • Running agents at scheduled times, and webhooks

There is also no way to move agents or records created in the standard version, so they need to be recreated in the self-hosted version. Pricing has not been published.

CoAI's take

In finance, healthcare and the public sector, AI agents' work records often contain confidential information and cannot be kept on outside services. The self-hosted version gives such organizations a way to use Claude's agents. Being able to choose AWS, Google Cloud or Microsoft's cloud, which they may already have contracts with, as the model provider also seems likely to make internal approval processes easier.

On the other hand, the operational work and responsibility increase, and some features are unavailable. If you do not need to keep data in-house, the easier standard version, or hosting only the sandbox yourself, is likely a better fit.

Sources

Tools mentioned in this article

Back to news