Skip to content
eastbaycyber

Disclosure: this page contains affiliate links. East Bay Cyber may earn a commission if you buy through them, at no extra cost to you. That never changes how we rank products. Editorial policy.

Self-hosted Claude Managed Agents keeps credentials in your environment

Breaking news 6 min read
EC
East Bay Cyber Editorial Team Updated Developing story
Built from the vendor or authority notice plus independent reporting; unconfirmed claims are labeled as reported. Human-reviewed after publishing.

TL;DR - Anthropic has made self-hosted Claude Managed Agents available in private beta. - Customers can run the harness, sandboxes, credentials, session history, server, database, and workers in their own cloud or on-premise environment. - Anthropic says credentials remain outside the sandbox, with a proxy adding them for MCP tools and supported API calls. - Customers can restrict network destinations and export agent activity over OpenTelemetry. - No CVE has been published for this announcement.

What happened#

Anthropic announced that self-hosted Claude Managed Agents is available in private beta. Managed Agents is described as a suite of composable APIs for building and deploying autonomous agents. The service provides the agent harness, sandboxes, credential handling, and session history for long-running agents.

With the self-hosted option, every part of the service runs in the customer’s cloud or on-premise environment. This includes the harness, sandboxes, credentials, and session history. Customers operate the Managed Agents server, its database, and the workers where agents run shell and file tools. Session history is stored in the customer’s own database.

Model requests can be sent to Anthropic directly, to Amazon Bedrock, Google’s Agent Platform, Microsoft Foundry, or a customer’s own gateway. Anthropic says zero data retention, when an organization has it in place with Anthropic, applies to requests sent to Claude in the same way as other Messages API calls. (Anthropic)

Anthropic presents the design around four principles. The harness, which calls Claude, is separated from the sandbox, where agent code runs. The agent loop, session history, and vault secrets remain outside the sandbox. Credentials also remain outside the sandbox. For MCP tools and supported API calls, a proxy outside the sandbox adds the credential to the request. Self-hosted customers can store credentials in their own secrets manager.

Network access can be limited to an allowlist of hosts. Anthropic says requests to other hosts are blocked. When customers host the sandbox themselves, they set the same limits in their own firewall. Managed Agents records every agent step, including tool calls and returned data, and exports the record over OpenTelemetry.

Are you affected#

Organizations are potentially affected if they are evaluating or operating self-hosted Claude Managed Agents in the private beta. Anthropic says access can be requested through an Anthropic account team, and documentation is available for getting started. (Anthropic)

The announcement does not publish a version number, CVE identifier, CVSS score, or exploitation status. It also does not describe a vulnerability in Claude Managed Agents. The change instead places more of the service’s operation inside the customer environment, including credential storage and handling, the server, database, workers, network restrictions, and session records.

Teams using agents with access to internal systems should identify whether they operate only a self-hosted sandbox or the full self-hosted Managed Agents service. Anthropic distinguishes those deployment models, noting that self-hosted Managed Agents runs the whole service in the customer environment.

What is confirmed and what is reported#

Claim Source Status
Self-hosted Claude Managed Agents is available in private beta. Anthropic Confirmed
Customers can run the harness, sandboxes, credentials, and session history in their own cloud or on-premise environment. Anthropic Confirmed
Customers operate the Managed Agents server, its database, and workers where agents run shell and file tools. Anthropic Confirmed
Credentials remain outside the sandbox, and a proxy adds them for MCP tools and supported API calls. Anthropic Confirmed
The customer can store credentials in its own secrets manager. Anthropic Confirmed
Network access can be restricted to allowed hosts, with other requests blocked. Anthropic Confirmed
Managed Agents records agent steps and exports them over OpenTelemetry. Anthropic Confirmed
The server, database, and firewall rules are the customer’s responsibility to run and configure. Help Net Security Reported
Anthropic limits credential proxy protection to MCP tools and supported API calls, so other services should be checked before deployment. Help Net Security Reported

What defenders should do now#

  1. Confirm whether your organization has access to the private beta and document which deployment model is in use: hosted Managed Agents, a self-hosted sandbox, or the full self-hosted service.

  2. Inventory the Managed Agents server, database, workers, credential stores, and session-history database operated in your environment.

  3. Follow Anthropic’s credential model by keeping credentials outside the sandbox. For MCP tools and supported API calls, use the proxy mechanism described by Anthropic. Where appropriate, store credentials in your own secrets manager. (Anthropic)

  4. Review every service an agent can reach. Anthropic’s credential proxy description covers MCP tools and supported API calls, so document how credentials are handled for other services before enabling access.

  5. Configure network restrictions using an allowlist of required hosts. Anthropic gives a git host and package registry as an example for an agent fixing broken builds. If the sandbox is self-hosted, apply the corresponding restrictions in the firewall. (Anthropic)

  6. Export agent records over OpenTelemetry and review tool calls and returned data alongside existing monitoring. The vendor says the record includes every step an agent took.

  7. Establish operational ownership for the customer-run server, database, workers, firewall rules, credential stores, and session history. Help Net Security reports that these components are now the customer’s to run and configure. (Help Net Security)

  8. Request access or consult Anthropic’s documentation through the Anthropic account team before deployment, as recommended in the announcement. (Anthropic)

This advisory is informational and is updated as vendors publish details. Confirm versions and actions against the vendor’s own guidance for your environment.

Sources

Last verified: 2026-11-13

Disclaimer: This article may contain affiliate links. We earn a commission on qualifying purchases at no extra cost to you.