Ollama serves a local model API on port 11434, but calling it from another machine fails when you are behind CGNAT or on a GPU/vast.ai host that blocks inbound ports. NeedPorts gives the Ollama endpoint a stable public address for your apps to call.
Ollama binds to a local port that is only reachable on the same machine or LAN by default. On a home connection behind CGNAT, or a rented GPU host that does not hand out public inbound ports, there is no public path to that port, so remote API calls cannot connect.
Call the Ollama API from an app, another server, or a laptop, using a stable URL that points at the GPU box actually running the model.
Requests to the host time out or refuse, because the GPU host or ISP does not expose an inbound port to the machine running Ollama.
NeedPorts gives you dedicated public ports over an outbound tunnel, so nothing depends on your ISP or router. After install, map one of your assigned ports to Ollama (local port 11434) and restart:
sudo needports use ollama <assigned-port>
sudo needports restartYour Ollama instance is then reachable at your dedicated public endpoint, for example your-server:30000, with no port forwarding on the local network.
Ollama does not authenticate callers by default, so an open endpoint lets anyone use your compute. Put an auth proxy or API key in front of it, restrict access where possible, and never publish it without some access control.
After mapping the Ollama port, send a request to the public endpoint from another machine (for example a small completion call) and confirm you get a model response. That proves the API is reachable through the tunnel.
NeedPorts plans start at $5/month or $30/year for 25 dedicated ports, which is plenty for Ollama and other self-hosted services on the same box.