Key Points
- Anthropic said Claude Managed Agents can now run entirely inside a customer's own cloud or data center.
- Previously Anthropic hosted the agent infrastructure, with self-hosting limited to sandbox environments only.
- The shift targets regulated firms barred by compliance rules from sending agent data outside their perimeter.
The latest:
Companies can now run Anthropic’s agent stack entirely on their own infrastructure, under a self-hosted deployment of Claude Managed Agents released in private beta. In a November 12, 2026 article, Anthropic said the agent harness, sandboxes, credentials and session logs all operate inside the customer environment. Access is granted by request through an Anthropic account team.
Details:
- What it is: Claude Managed Agents is Anthropic’s set of APIs for building autonomous agents. The self-hosted option runs that service inside the customer’s own private cloud or on-premises infrastructure, rather than on Anthropic’s systems, Anthropic said in its article announcing the private beta.
- What moves in-house: According to Anthropic, the customer runs the Managed Agents server, its database, and the worker machines where agents execute shell and file tools. Previously Anthropic hosted that infrastructure itself, and self-hosting covered sandbox environments only.
- Credential handling: Vaults are designed to keep credentials outside the sandbox, Anthropic said. For supported MCP tools and API calls, a proxy sitting outside the sandbox injects the credentials into the request. Self-hosting customers can store credentials in their own secrets manager.
- Session data: Session logs are stored in the customer’s database, according to Anthropic. Organizations holding a zero data retention agreement with the company are covered for requests sent to Claude on the same terms that apply to other Messages API calls.
- Security design: Anthropic listed four principles: separating the agent harness from the sandbox, keeping credentials out of the sandbox, restricting network access to an allowlist of hosts, and maintaining a complete exportable record of every session through OpenTelemetry.
- Model routing: Customers choose where inference requests go, Anthropic said. The options are Anthropic directly, Amazon Bedrock, Google Agent Platform, Microsoft Foundry, or the organization’s own gateway — leaving the model call itself under customer-controlled routing.
- Who gets in: The private beta is aimed at developers building agents on the same core APIs that power the hosted service, according to Anthropic. The company said access requires a request: “Contact your Anthropic account team to request access.”
- Pricing unstated: Anthropic did not publish pricing for the self-hosted version in the announcement, directing readers instead to its pricing page and sales team. The company also did not name a date for general availability beyond the private beta.
Background:
Claude Managed Agents gives developers prebuilt infrastructure for running autonomous agents — sandboxed machines, tool access and session tracking — instead of assembling it themselves. Until this release, that infrastructure ran on Anthropic’s side, with only the execution sandboxes available for customer hosting.
Between the lines:
The design choices point at a specific buyer. Keeping session logs in the customer’s database, honoring zero data retention, allowing credentials in the customer’s own secrets manager, and permitting model calls to route through Bedrock, Google, Microsoft or a private gateway together remove the usual compliance objections to agents — namely that data, secrets and logs leave the perimeter. The OpenTelemetry export addresses the audit requirement that typically follows.
What’s next
Watch for Anthropic to publish pricing for the self-hosted tier and to set a general availability date, plus any expansion of which MCP tools and API calls support proxy-injected credentials.
