← All articles
Anthropic lets companies run AI agents on their own servers
AI News

Anthropic lets companies run AI agents on their own servers

Anthropic launched Self-hosted Claude Managed Agents in private beta, so AI agents, credentials and session history can stay inside a company's own infrastructure.

AI agents have started doing important work without supervision: fixing broken builds, testing releases and triaging internal alerts. Until now, all of their records and credentials lived on Anthropic's servers. On 12 November 2026, Anthropic opened a new option called Self-hosted Claude Managed Agents in private beta, letting companies run the whole service in their own cloud or on their own hardware.

1. What was launched

Managed Agents is a set of APIs for building and running autonomous AI agents. It provides four things: the agent harness, the sandbox where agent code runs, credential handling, and session history.

From 12 November, all four can run inside a company's own environment. This is a private beta: companies request access through their Anthropic account team. No pricing has been announced.

2. What the customer runs

In the self-hosted setup, the customer operates three components: the Managed Agents server, a PostgreSQL database that stores session history, and the workers where agents run their shell and file tools.

The Claude model itself still runs outside the customer's network. Model requests go wherever the customer points them: directly to Anthropic, to Amazon Bedrock, Google's Agent Platform, Microsoft Foundry, or the company's own gateway. Developers use the same core APIs as the hosted version, so existing agents do not need a rewrite.

3. Security built into the system

Anthropic argues that an unattended agent cannot rely on a person checking every step, so the limits have to be built into the system itself. It uses four principles:

  1. The harness, which calls Claude, is kept separate from the sandbox where agent code runs.
  2. Credentials stay out of the sandbox. A proxy outside the sandbox attaches the credential to each request, and self-hosted customers can keep credentials in their own secrets manager.
  3. Each environment can be locked to a list of allowed hosts. Requests to any other host are blocked.
  4. Every step an agent takes is logged, including tool calls and their results, and the log exports over OpenTelemetry into existing monitoring tools.
Official Anthropic illustration for Self-hosted Claude Managed Agents, a laptop with a keyhole symbol
Anthropic's official illustration for Self-hosted Claude Managed Agents, highlighting security. Image: Anthropic.

4. Why it matters

Banks, hospitals and defense companies have hesitated to allow AI agents because data and credentials would have to leave their perimeter. With self-hosting, session history and credentials stay on their own servers, while model requests remain covered by a zero data retention agreement if one is in place.

Anthropic launched Managed Agents as a hosted service in April, then added self-hosted sandboxes in May. This release completes the set: fully hosted, customer-hosted sandboxes, and now fully self-hosted.

The New Stack also highlighted a detail in the software license: if Anthropic flags a security risk in software a customer runs, the customer has eight hours to act. Even on the customer's own servers, Anthropic keeps some control.

What it means for you

If you work at a company with strict data rules, this is more than another product launch.

  1. Large companies have a new option. IT teams that previously rejected AI agents over data concerns can evaluate this beta through their Anthropic account team.
  2. Operational costs shift to you. The server, database and firewall rules are now the customer's responsibility to patch and maintain.
  3. No pricing yet. Do not budget for it; Anthropic has not announced pricing for the self-hosted option.
  4. Vendor control remains. The eight-hour response requirement for flagged security risks means you should read the license terms before signing.
  5. Regular users are unaffected. Everyday Claude users need to do nothing; this is for companies running autonomous agents on internal systems.

Sources