Feature proposal: first-class "bring your own mesh" (ZTNA) for MCP + Remote Control — from a phone↔PC mirror-bridge field setup

Status Open
Maintainer reply None cached
Activity 0 comments · opened Aug 17, 2026

Feature proposal: first-class "bring your own mesh" (ZTNA) for MCP + Remote Control

From field experience: I run an always-on personal Linux workstation as my working environment, with a phone↔PC mirror-bridge: multiple input surfaces (terminal, phone, watch) converge on the same live machine over a consumer mesh VPN, with Claude Code Remote Control layered on top. The result is dual simultaneous control — local sessions keep working while I drive the same environment from my phone — and it works well in daily use. But every piece had to be glued together by hand.

The gap this exposes

  1. Enterprise MCP reachability. An organization's most valuable MCP servers sit on private networks (internal databases, on-prem systems). Today, reaching them from Claude requires a publicly reachable endpoint — exactly what security teams refuse. In my experience this is the single biggest deployment blocker for enterprise MCP.
  2. Heavy streams. Relay-only transport is a poor fit for the multimodal remote experiences that are coming (screen mirroring, voice, camera).

The proposal

Not "Anthropic operates a WireGuard network" — running a global mesh is its own business and liability. Instead: first-class integration with the zero-trust overlays organizations (and power users) already run — Tailscale, Cloudflare, and similar:

  • an org (or user) declares its overlay/tailnet once, in the console;
  • MCP connectors and Remote Control endpoints resolve inside the overlay — nothing exposed to the public internet, per-tenant isolation by construction;
  • direct peer-to-peer (E2E-encrypted) paths for heavy streams, with relay fallback;
  • a privacy argument that sells itself: E2E P2P means less user data transiting Anthropic relays, not more.

One observation from running dual control daily

The moment two hands (or a human and an agent) share one live environment, the first question any regulated organization will ask is: "who commanded what, and how do you prove it?" Multi-surface control creates natural demand for tamper-evident, signed session/audit trails as a first-class feature.

My setup is the manual existence proof that the pattern works. The proposal is simply: make the glue native.

---
Roberto Locatelli (roberto.locatelli.81@gmail.com)

View original on GitHub ↗