Anthropic opens Claude Managed Agents to self-hosted sandboxes

Anthropic launches Claude Managed Agents in self-hosted sandboxes with MCP tunnels, allowing secure tool execution inside private enterprise networks.

Anthropic is evolving its Claude Managed Agents: an agent can now run in an enterprise-controlled sandbox and connect to its private MCP servers, all within the customer's security perimeter. On the Claude Platform, the self-hosted sandbox is entering public beta, while MCP tunnels remain in early access by request.

With the self-hosted sandbox, sensitive files, dependencies, and services remain on the enterprise's infrastructure, or with a managed provider like Cloudflare, Daytona, Modal, or Vercel. The agent loop, which manages orchestration, context, and error recovery, remains hosted by Anthropic, but tool execution shifts to the customer's environment, with its network rules, audit logs, and resource sizing. Enough to absorb heavy workloads, such as long compilations or image generation.

MCP tunnels, for their part, allow agents to reach MCP servers located in a private network without exposing them to the public Internet: internal databases, private APIs, or ticketing tools become callable. A lightweight gateway establishes a single outbound connection, without inbound firewall rules or a public access point, with end-to-end encrypted traffic. The feature is supported in Managed Agents and the Messages API, and is controlled from the Claude console. Several customers, including Amplitude, Clay, Rogo, and DoorDash, are already building agents on these building blocks.