Skip to content
Neural Daily

ToolsNews

Self-hosted Claude Managed Agents enters private beta

🇭🇺 Olvasd el magyarul →

Anthropic opened a private beta of self-hosted Claude Managed Agents. The full agent stack, meaning the server, sandboxes, credentials and session history, runs in your own cloud or on-premises, and model requests go wherever you point them.

· 2 min read

Rajzolt laptop kulcslyuk ikonnal, a self-hosted Claude Managed Agents bejelentő képe
Image: Anthropic · Press kit · Source

What changed

Managed Agents is Anthropic's suite of APIs for running agents: harness, sandboxes, credential handling and session history for long-running agents. Until now it ran on Anthropic's servers, and since May only the sandboxes could move elsewhere. Now the whole service can run on your own infrastructure (Anthropic).

  • You operate the Managed Agents server, its PostgreSQL database, and the workers where agents run their shell and file tools.
  • Session history stays in your own database.
  • Model requests can go to Anthropic directly, Amazon Bedrock, Google's Agent Platform, Microsoft Foundry or your own gateway.
  • Developers build agents with the same core APIs as the hosted service.

The security model

According to the docs, Anthropic has no remote access to the server or its log, and the server connects to Anthropic only when the Claude API is its inference provider (Claude docs). Every credential the server accepts is one you set, one your identity provider issued, or one the server generated itself, such as the token it hands a worker.

The harness and the sandbox are separated: the agent loop, session history and vault secrets stay outside the sandbox. For MCP tools and supported API calls, a proxy outside the sandbox attaches the credential. Anything else an agent reaches is worth reviewing before go-live.

How to try it

The quickstart ships a Docker Compose stack: the server, PostgreSQL and workers, with each session in its own sandbox container (Claude docs). It is for development only and serves plain HTTP. For production Anthropic recommends two Helm charts on Kubernetes. Access goes through your Anthropic account team; pricing and a GA date were not disclosed.

Why it matters

In regulated environments where code, internal systems and session logs cannot leave the network, this was the main blocker. Note that Claude itself still runs outside your network: prompts go out and completions come back, with ZDR applying as for any Messages API call. Competition among managed agent platforms is heating up, with Mistral Studio moving in a similar direction (Mistral Studio: managed agents, workflows and apps).

    By Mikola Máté Patrik

    Produced with the help of AI, reviewed by an editor. How we make our articles

    Reviewed and edited by the Neural Daily desk. About

    Related articles