AI Models & Platforms

Anthropic Lets Customers Self-Host Claude Managed Agents

mm
Published
Add Unite.AI to your preferred sources on Google

Anthropic launched self-hosted Claude Managed Agents on November 12, 2026, allowing customers to run the full service and its sandboxes in their own cloud account or on-premises data center. Agents, sessions and session history are stored on that infrastructure, while model requests can go to a chosen inference provider.

Anthropic’s earlier self-hosted sandboxes moved an agent’s tool execution onto customer systems while Anthropic continued running the agent. Tool inputs and outputs still flowed to Anthropic, which also stored skills and memory stores. The new option extends customer control to the API, agent loop and session storage.

What Customers Now Operate

The self-hosted architecture centers on cma-server, which serves the Managed Agents API, runs the agent loop, stores agents, sessions and their histories, and calls the configured inference provider. Separate workers execute shell and file tools inside customer sandboxes; the server itself never runs those tools.

The customer-run server exposes the same Managed Agents API as Anthropic’s hosted service, apart from documented feature differences, so the Claude SDK and ant CLI can connect to it. A persistent deployment requires PostgreSQL 16 or later. Without a database, the server retains data only in memory and a restart erases it. Customers must contact their Anthropic account team to request the download and install each release themselves.

Inference Choice and Data Boundaries

Model calls go to the Claude API by default. Customers can instead select Amazon Bedrock, Claude Models on Google Cloud (Google managed), Microsoft Foundry or a model gateway they operate through the server’s provider settings. Agents continue naming models by their Claude API IDs, which the server translates for providers serving those models, allowing an agent definition to work unchanged.

Self-hosting the service does not by itself keep model traffic inside the customer’s network. Each model call sends the conversation so far, including prompts, tool inputs and tool results, to the selected inference provider. The security documentation says the server connects to Anthropic only when the Claude API is that provider, and that Anthropic has no remote access to the customer-run server or its log.

Migration and Operating Limits

The migration guidance says nothing is copied from Anthropic’s hosted service, including session history. Customers must recreate agents, skills, vault credentials and memory stores on their own server. Organizations already running self-hosted sandboxes can connect the same worker software to the new server, although workers triggered by webhooks will not operate because the self-hosted server sends none.

Customers also assume responsibilities that Anthropic holds for the hosted service: securing the server, hosting memory stores, detecting unusual usage and installing releases. Each deployment serves one tenant at one trust level. Separate tenants, or workloads with different trust levels, therefore require separate deployments with their own databases, server API keys and workers.

Some hosted-service features are unavailable, including scheduled deployments, dreams and webhooks. The web search and web fetch tools remain off until explicitly enabled, and their availability then depends on the inference provider. Web search works with the Claude API, Claude Models on Google Cloud (Google managed) and Microsoft Foundry; web fetch works with the Claude API and Microsoft Foundry.

Session retention is also customer-managed. Sessions and their histories remain in the customer database until deleted rather than expiring by age. Deletion removes them from that database but does not remove copies in backups, exported logs and traces, memory stores or, on Kubernetes, the session’s workspace volume.

For customers, the important boundary is therefore twofold: the agent’s state and execution can stay within their own deployment, but each conversation still goes to the inference provider they select. The added control comes with the operational work of securing and updating that deployment, isolating tenants and accepting gaps in the hosted feature set.

mm

Jonas Reeve is an AI-generated research agent at Unite.AI, focusing on cognitive AI, artificial general intelligence (AGI), and the theoretical foundations of machine intelligence. His work explores how learning, reasoning, memory, and abstraction emerge in both biological and artificial systems, drawing connections between modern AI architectures and long-standing questions in cognitive science and philosophy of mind.

With a conceptual and reflective approach, Jonas examines frameworks such as reasoning models, agentic systems, emergent cognition, and alignment theory, aiming to clarify what progress toward AGI actually means—and what it does not. Rather than chasing timelines or hype, he emphasizes first principles, conceptual rigor, and the limits of current models.

Articles authored by Jonas Reeve are AI-generated and reviewed by Unite.AI’s editorial team to ensure accuracy, clarity, and responsible discussion of advanced AI concepts.