Setting OLLAMA_HOST=0.0.0.0 gets Ollama answering on your LAN in about ten seconds. Reaching that same endpoint from a laptop on a different network, a phone on mobile data, or a VPS running Cursor is a different problem, and it's the one most people actually have.
The usual next steps all work, and all cost you something:
- Port forwarding means a router login, a static rule that breaks the next time your ISP renews your lease, and an open inbound port sitting on a machine that's also running an LLM.
- ngrok or a Cloudflare Tunnel gets you a public URL fast, but now you're proxying through a third party's infrastructure and, for ngrok specifically, usually paying once you want a stable subdomain.
- Behind CGNAT, which covers most mobile connections and a growing share of home connections, port forwarding isn't even on the table. There's no public IP to forward to.
Amallo skips the whole category of problem by not listening for inbound connections at all. It runs in the menu bar or system tray, holds one outbound connection open to Relay, and hands you an OpenAI-compatible base URL and key. Nothing is exposed on your router, because nothing on your router needs to change.
- Have Ollama running locally. Amallo expects it on
127.0.0.1:11434 with at least one model already pulled. It doesn't start Ollama for you, it just finds it there. - Install Amallo for macOS or Windows. It opens with a base URL and a
41t_… key already generated, no account, no signup. - Point a client at it. Paste the base URL and key into Cursor, Aider, Open WebUI, or anything else that speaks the OpenAI API. Streaming works out of the box.
- Leave it running. The connection is outbound-only and stays parked at Relay, so it survives normal NAT and CGNAT without any configuration on your end.
That's the whole setup. No ssh -R, no cloudflared tunnel, no router admin panel.