Anthropic Will Now Let Companies Run the Entire Claude Managed Agents Service Inside Their Own Network

Anthropic opened a private beta of self-hosted Claude Managed Agents, letting organisations run the agent server, PostgreSQL session store and sandbox workers in their own cloud or data centre while sending model calls to Anthropic, Amazon Bedrock, Google Cloud, Microsoft Foundry or their own gateway.

Small robot helpers tend emerald plant beds and blush-pink flowers inside a walled pewter-stone greenhouse.
Small robot helpers tend emerald plant beds and blush-pink flowers inside a walled pewter-stone greenhouse.

Anthropic has started letting companies host the whole of Claude Managed Agents themselves, not just the sandboxes where agents run code. Self-hosted Claude Managed Agents went into private beta this week, with access granted through a customer’s Anthropic account team. Organisations run the agent harness, the sandboxes, the credentials and the session history in their own cloud account or on-premise data centre, and choose where requests to Claude are sent.

From hosted service to full self-hosting

Managed Agents is Anthropic’s set of composable APIs for building and deploying autonomous agents. It supplies the agent loop, sandboxes, credential handling and session history needed to put long-running agents into production. Anthropic first offered it as a hosted service on the Claude Platform, then added self-hosted sandboxes for teams whose code had to stay in their own environment.

In its announcement, Anthropic says agents are now taking on more critical work, such as fixing broken builds, testing releases, triaging alerts and working through bug backlogs, and that they run unattended with access to internal systems. For that level of access, some teams want the entire service in their own environment, which until now meant building the platform themselves.

The documentation lays out the three options plainly. With cloud sandboxes, Anthropic runs everything. With self-hosted sandboxes, the customer hosts sandboxes while Anthropic runs the service and the model. With self-hosted Managed Agents, the customer hosts both the sandboxes and the service, and the model runs at Anthropic or at a provider the customer picks.

A server, a Postgres database and your own workers

A deployment is made of one or more servers, a database and the customer’s sandboxes:

  • cma-server serves the API, runs the agent loop and stores agents, sessions and history. It calls the inference provider when the agent needs the model, and never runs the agent’s shell or file tools itself.
  • PostgreSQL 16 or later stores the data. Without a database, the server keeps everything in memory and a restart erases it.
  • Workers, run as an ant beta:worker command or an SDK worker, poll the server for work and execute shell and file tools inside the customer’s sandboxes.
  • A web console on the server lets people sign in through the customer’s identity provider, and an admin CLI, cma-server-admin, creates environments and their keys.

Because the self-hosted server exposes the same API as the hosted one, the SDK and the ant CLI work against it unchanged. Model calls can go to the Claude API, Amazon Bedrock, Claude models on Google Cloud, Microsoft Foundry or a custom gateway. The server software is licensed under a dedicated Claude Managed Agents Server agreement rather than an open-source licence.

The security model shifts to the customer

Anthropic frames the product around four design principles that apply however Managed Agents is hosted. The harness, or “brain,” is kept separate from the sandbox, or “hands,” so session history and vault secrets stay outside the place where agent code runs. Credentials are added by a proxy outside the sandbox for MCP tools and supported API calls, and self-hosters can keep them in their own secrets manager. Each environment can be limited to an allowlist of hosts. Every session leaves a full record of tool calls and results, exportable over OpenTelemetry.

The documentation is direct about what customers take on. The agent’s bash tool is a real shell with the worker’s permissions and network access, and the worker does not isolate it, so isolation comes from the container or VM it runs in. The supplied Docker Compose stack and Helm chart give each session its own container or pod. Each deployment serves one tenant at one trust level; separate tenants or trust levels need separate deployments with their own database, API key and workers.

What still leaves the network

Anthropic says it has no remote access to the server or its logs, and the server connects to Anthropic only when the Claude API is its inference provider. In that case, each model call carries the conversation so far plus three headers holding session, agent and environment IDs, which cannot be turned off. Requests also carry a User-Agent naming the server software, and some calls carry a cma-server label that Anthropic can use to count self-hosted usage. Session history stays in the customer’s database until deleted, and existing zero data retention arrangements apply to model calls as they do for any other Messages API request.

Anthropic has not published pricing for the self-hosted option or a date for general availability.

Topics
  • #Anthropic
  • #Claude
  • #Managed Agents
  • #Enterprise AI
  • #Agent Infrastructure
Raj M

Author

Raj M

Contributor

AI Systems Architect is a seasoned technology leader with over 15 years of experience in the IT industry working with Fortune 500 companies. With a solid foundation in multi-agent systems, open-source LLM infrastructure, and enterprise deployment, he excels at building scalable production-grade AI platforms.