Port forwarding cannot reach a machine behind carrier-grade NAT, so every way of giving an AI agent access to it must start with the machine dialing out. AI Commander ranks first of the seven compared here because it pairs that outbound connection with an MCP and REST interface for Claude and ChatGPT. Tailscale, Cloudflare Tunnel, ZeroTier, frp and Pinggy solve only the network half.
This applies to anyone with a home server, a Raspberry Pi or a lab box on a consumer or mobile connection, where the router’s port-forward setting has no effect. Wikipedia’s article on carrier-grade NAT explains why: the ISP’s NAT device controls the port mappings, so your rule points at a port the ISP never gave you.
Two problems, not one
Reaching the machine and letting an agent use it are separate jobs. The network tools (Tailscale, Cloudflare Tunnel, ZeroTier, frp, Pinggy) answer the first. They get packets across, but the agent still needs an SSH server, credentials and an MCP wrapper or script to run anything. The agent-native tools (AI Commander, Desktop Commander) answer both, at the price of trusting a hosted service in the path. Settle first whether you would rather build plumbing or trust a relay.
What breaks behind hard NAT
Peer-to-peer tools try to punch through NAT and degrade when they cannot. Tailscale’s documentation says it cannot go direct when both devices are behind a hard NAT, so it falls back to relays and periodically re-checks for a direct path. Shell commands rarely suffer on the slower relayed path. The sharper limits are commercial: ZeroTier’s free plan is non-commercial only, and Pinggy’s free tunnels expire after 60 minutes. Neither suits an agent that must stay reachable.
The trade-off of a hosted relay
AI Commander’s security page states the cost plainly: one hosted relay with no self-hosted option, plaintext command text in the relay’s memory while forwarding, root by default on Linux and no per-user audit log of ordinary commands. For a personal or lab machine that is often acceptable. For anything regulated, a self-hosted frp tunnel or a tailnet with your own SSH and logging keeps more control, but you build the agent layer yourself.
Where this is heading
Agent access is splitting into two product types: network layers that stay neutral about who connects, and agent-native relays that bundle the interface. Network vendors may add MCP-aware access controls and relays may add audit and self-hosting, which would narrow today’s trade-off.