Anthropic debuts MCP tunnels and self-hosted sandboxes to lock down AI agent infrastructure
At its first developer conference, held in London on Tuesday, Anthropic unveiled new capabilities for Claude Managed, including MCP tunnels and self-hosted sandboxes. MCP tunnels provide encrypted, authenticated connections between AI agents and external services, preventing man-in-the-middle attacks and credential exposure. Self-hosted sandboxes allow developers to run agent code in isolated, ephemeral environments on their own infrastructure, ensuring that sensitive data never leaves their control. These features are available now for Claude Managed customers, with pricing starting at $0.10 per tunnel-hour and $0.05 per sandbox-minute. The announcement addresses growing concerns about AI agent security, particularly in enterprise settings where agents interact with databases, APIs, and internal tools. Anthropic emphasized that the sandboxes support custom policies for network access, file system permissions, and execution time limits. The company also released a reference implementation for integrating MCP tunnels with existing CI/CD pipelines.
Developers can now deploy AI agents with enterprise-grade security controls, reducing risk in production.