Securing AI agent credentials with MCP tunnels

Anthropic added self-hosted sandboxes in public beta and MCP tunnels in research preview for Claude Managed Agents.

The new setup separates agent orchestration from tool execution, keeping credentials at the network boundary instead of inside the agent.

This lowers the chance that a compromised agent can reach internal systems through exposed tokens.

It also gives enterprises more control over where tools run and how agents connect to private services.