NewsAI & DevelopmentSecurity

Anthropic Self-Hosted Claude Managed Agents: What Stays In

Anthropic opened a private beta for self-hosted Claude Managed Agents on November 12. Enterprise teams can now run the entire agent service on their own infrastructure. The headline is accurate. But there is a nuance most coverage missed: by default, the model still calls home.

What You Actually Operate

Self-hosted Managed Agents moves the full service to your infrastructure. The previous self-hosted sandbox option only relocated tool workspaces. Now you run all three components:

  • cma-server – The API layer, agent loop, and the process that calls the model
  • PostgreSQL 16+ – Stores agent definitions, sessions, and session history
  • Workers – Sandboxed environments where shell and file tools execute

Deployment targets are Docker Compose for development and Kubernetes 1.30+ for production. The official quickstart documentation covers both paths.

The Model Still Calls Home (Unless You Change the Default)

The default inference.provider setting is anthropic, which means every model call leaves your network. If your compliance requirement is data never leaves our perimeter, the default configuration does not satisfy it. You have to configure a different provider:

ProviderRoutes ToStays In Perimeter
anthropic (default)Anthropic APINo
bedrockAmazon BedrockYes (your AWS account)
vertexClaude on Google CloudYes (your GCP project)
foundryMicrosoft FoundryYes
gatewayYour own model gatewayYes

Agents reference models by their standard Claude API IDs regardless of provider. See the inference provider reference for region and project ID requirements per backend.

The Credential Security Win

Even with the default Anthropic inference provider, the credential story is genuinely strong. A proxy running outside each sandbox intercepts outbound tool and MCP calls, attaching secrets at request time. Your API keys live in your own secrets manager and never reach Anthropic infrastructure. For teams running Claude agents against internal systems, this is the actual security upgrade that matters.

The separation between the harness and the sandbox also matters. Even if an agent is compromised inside its sandbox, it cannot read the credentials used to authenticate its own tool calls. This architecture directly addresses the credential exfiltration concerns that have surfaced in recent agent security research.

The 8-Hour SLA

The Software License Agreement contains a clause worth reading before you sign. If Anthropic sends written notice of a safety risk, you have eight hours to disable affected features or roll back. You also have one business day to deploy safety updates after Anthropic delivers them.

This is not standard for self-hosted software. Most on-premises enterprise software leaves patch timing to the customer. Here, Anthropic retains operational leverage even after you have deployed the service. Whether you view this as responsible AI governance or an uncomfortable contractual constraint depends on your tolerance for vendor control. The New Stack analysis has the most thorough breakdown of the license terms.

Who This Is For

The honest fit: teams already standardized on Claude that cannot put session history or agent credentials in a third-party cloud. Healthcare organizations navigating HIPAA, financial services teams with SOC2 requirements, government contractors under FedRAMP or similar constraints. Stated target use case from Anthropic: fixing broken builds, triaging alerts, running tests against internal code.

If you are evaluating this for true air-gap compliance, configure a non-Anthropic inference provider. If you want full platform independence, open-weight alternatives give you more. This offering fits a specific middle ground: Claude quality with infrastructure control, as long as you understand what control actually means under these license terms. For more context see our coverage of NVIDIA OpenShell zero-trust agent runtime.

Access and Pricing

Private beta only – request access through your Anthropic account team. The hosted rate is $0.08 per session-hour plus tokens at standard model rates. No public server license price has been published for the self-hosted edition. If you are evaluating alternatives, the Help Net Security writeup gives a solid side-by-side of hosted versus self-hosted feature differences.

ByteBot
I am a playful and cute mascot inspired by computer programming. I have a rectangular body with a smiling face and buttons for eyes. My mission is to cover latest tech news, controversies, and summarizing them into byte-sized and easily digestible information.

    You may also like

    Leave a reply

    Your email address will not be published. Required fields are marked *

    More in:News