Anthropic Opens Self-Hosted Claude Managed Agents in Private Beta

By Hector Herrera | November 14, 2026 | Security

Anthropic launched a private beta that lets enterprises run the full Claude Managed Agents stack — the orchestration server, tool sandboxes, and session-history database — inside their own cloud environment or on-premises infrastructure. Only the model inference call leaves the customer's perimeter. Everything else stays put.

The move addresses the most common blocker Anthropic hears from regulated industries: data leaves the building. Healthcare systems, financial institutions, and defense contractors routinely pass on AI agent deployments because session histories, tool call logs, and in-flight data flow through third-party servers. Self-hosted managed agents are a direct answer to that objection.

What "Self-Hosted" Actually Means Here

There is a spectrum of what vendors mean by self-hosting. At the shallow end: a hosted service with a private subdomain. At the deep end: the full stack running in your data center, air-gapped from the internet.

Anthropic's implementation sits closer to the deep end. According to the Unite.AI report, customers who enroll in the beta deploy:

• The agent orchestration server — which routes tasks, manages tool calls, and handles multi-step reasoning
• Tool sandboxes — isolated execution environments for the code, file access, and external API calls agents make
• Session-history database — the memory store that lets agents maintain context across conversations

The only traffic that leaves the customer environment is the inference request to Claude itself. Anthropic still hosts the model; the customer controls everything around it.

Who This Is For

The private beta is not for general availability. Customers must contact their Anthropic account team to request access, which signals this is initially targeting mid-to-large enterprise accounts with dedicated sales relationships — companies already running Claude at scale who need a compliance upgrade path, not net-new customers.

The industries where this matters most:

• Healthcare: HIPAA requires that protected health information not leave covered entity control except under specific agreements. AI agents that process patient records generate session data that could be considered PHI. Self-hosting eliminates that exposure.
• Finance: SOC 2 and financial regulators increasingly scrutinize where data flows during AI-assisted decisions. Banks using Claude for loan analysis or trading compliance want logs that never leave their environment.
• Government and defense: FedRAMP, ITAR, and similar frameworks restrict data handling in ways that most cloud AI services cannot satisfy. Self-hosting is a prerequisite for many federal deployments.

What This Means for the Enterprise AI Market

Anthropic is not the first to offer some form of self-deployment — Microsoft Azure OpenAI, Google Vertex AI, and AWS Bedrock all offer private deployment options for foundation models. What is different here is that Anthropic is also self-hosting the agentic layer, not just the model weights.

Agents are stateful. They remember. They call tools. They write logs. The compliance risk of an agent is not just what model it runs on — it is every intermediate state between the user's input and the final answer. By including orchestration and session history in the self-hosted package, Anthropic is addressing the full audit surface.

For competitors, this raises the bar. Enterprises evaluating agent platforms will now ask: can I run the whole thing in my environment? Not just the model, but the agent runtime.

What to Watch

The private beta will reveal how much demand exists for a fully self-hosted agentic stack at Claude's inference cost. Watch for Anthropic to either expand beta access broadly or announce a GA date with pricing — the latter will indicate how central self-hosted managed agents are to Anthropic's enterprise revenue strategy.

Source: Unite.AI, 2026