Anthropic Opens Claude Managed Agents to Self-Hosting

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 accounts or on-premises data centers.

Anthropic’s release moves three control points to customer infrastructure: the Managed Agents API, the agent loop and session storage. The company said agents, sessions and session histories now stay on customer systems, while model requests can go to a chosen inference provider.

The option extends Anthropic’s earlier self-hosted sandbox arrangement, which moved tool execution to customer systems while the company continued running the agent. Under that earlier setup, tool inputs and outputs still flowed to Anthropic, which also stored skills and memory stores.

Anthropic’s security documentation puts a boundary around that control. It says: “the server connects to Anthropic only when the Claude API is that provider.” The documentation also says Anthropic has no remote access to the customer-run server or its logs.

What changes when customers run the service?

Anthropic’s self-hosted architecture centers on cma-server. According to the company, cma-server:

• serves the Managed Agents API;
• runs the agent loop;
• stores agents, sessions and session histories; and
• calls the configured inference provider.

Separate workers execute shell and file tools inside customer sandboxes, while the server itself never runs those tools. Anthropic said the customer-run server exposes the same Managed Agents API as its hosted service, apart from documented feature differences. The Claude SDK and ant CLI can connect to it.

A persistent deployment requires PostgreSQL 16 or later. Anthropic said a server without a database 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.

Anthropic’s design gives customers the agent runtime and stored state while leaving tool execution in customer-controlled sandboxes.

Which inference providers can customers use?

Anthropic routes model calls 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.

Anthropic said agents continue naming models by their Claude API IDs. The cma-server translates those IDs for providers that serve those models, allowing an agent definition to work unchanged.

Anthropic warned that 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.

Provider choice changes the inference destination, not the operating duties that come with running cma-server and its workers.

What do customers have to migrate and operate?

Anthropic’s 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.

Anthropic said organizations already running self-hosted sandboxes can connect the same worker software to the new server. Workers triggered by webhooks will not operate because the self-hosted server sends no webhooks.

Anthropic said customers assume responsibilities that Anthropic handles for the hosted service: securing the server, hosting memory stores, detecting unusual usage and installing releases.

Anthropic limits each deployment to one tenant at one trust level. Separate tenants, or workloads with different trust levels, require separate deployments with their own databases, server API keys and workers.

Self-hosting starts with manual reconstruction and adds a standing infrastructure-management load.

Which hosted features are missing?

Anthropic identifies three hosted-service features unavailable in the self-hosted version:

• Scheduled deployments.
• Dreams.
• Webhooks.

Anthropic said web search and web fetch remain off until a customer explicitly enables them. Availability then depends on the inference provider:

• Web search: Claude API, Claude Models on Google Cloud (Google managed) and Microsoft Foundry.
• Web fetch: Claude API and Microsoft Foundry.

The hosted and self-hosted products are not feature-for-feature equivalents.

How long does customer-managed session data remain?

Anthropic says sessions and their histories remain in the customer database until deleted rather than expiring by age. Deletion removes them from that database.

Anthropic’s deletion guidance also says copies can remain in backups, exported logs and traces, memory stores or, on Kubernetes, the session’s workspace volume. Deleting a session from the live database does not erase every copy of its data.

Anthropic describes a twofold boundary: an agent’s state and execution can remain inside the customer deployment, but each model call still reaches the selected inference provider. Customers gain control over cma-server, sandboxes, session storage and retention, but they also take on security, updates, usage monitoring and tenant isolation.

Based on Anthropic’s documented limits, adoption will hinge on whether customers can absorb those duties while retaining Claude SDK and ant CLI compatibility, and whether missing scheduled deployments, dreams, webhooks and provider-specific web tools prove tolerable.