# Anthropic MCP Tunnels: Agents Reach On-Prem Data Privately

> Anthropic's MCP tunnels let a managed agent act on data living on your own server through an MCP connection, so sensitive data never leaves your infrastructure.

Published: 2026-05-22
URL: https://daniliants.com/insights/claude-update-mcp-tunnels/
Tags: mcp, claude, ai-agents, anthropic, self-hosting, data-privacy

---

## Summary

A short social clip recapping Anthropic's "MCP tunnels" announcement at the Code with Claude conference in London. The feature lets a managed agent reach data living on your own server through an MCP connection, instead of pulling that data out to an external service. The pitch: keep sensitive data on-prem while still letting agents act on it.

## Key Insight

- **MCP tunnels = agent reaches your private backend without exfiltrating data.** The agent accesses website/server data through an MCP endpoint that stays on your infrastructure.
- **Framing distinction worth keeping:** humans use the front end (click around), agents use the back end (MCP). MCP is the agent-facing API surface of a system.
- **The real value is data residency**, not new capability per se, work happens against data that never leaves your server. Relevant for compliance-sensitive or client-confidential workloads.
- Source is a hype-style social clip, so treat specifics (exact mechanics, GA vs preview) as unverified, confirm against Anthropic's actual docs before building on it.