Why Ollama says connection refused, and how to fix it

Ollama runs as a background service (a systemd unit on Linux, a menu bar app on macOS, or a tray app on Windows) that listens on port 11434. If the process crashed, was never started, or was stopped by a system update, there is nothing listening on that port and any connection attempt is refused immediately at the TCP level.

Fix: Start the service directly with ollama serve in a terminal to see any startup errors, or restart the managed service with systemctl restart ollama on Linux. On macOS or Windows, quit and relaunch the Ollama app from the applications menu.

systemctl status ollama
systemctl restart ollama
ollama serve

By default Ollama binds to 127.0.0.1, which only accepts connections originating from the same machine. A request from a Docker container, a different VM, or another host on the network is refused even though the service is healthy, because the kernel never routes that traffic to a loopback-bound socket.

Fix: Set the environment variable OLLAMA_HOST=0.0.0.0:11434 before starting the service so it binds to all interfaces, then restart Ollama. Pair this with a firewall rule or reverse proxy since it removes the implicit localhost-only protection.

export OLLAMA_HOST=0.0.0.0:11434
systemctl set-environment OLLAMA_HOST=0.0.0.0:11434
systemctl restart ollama

Even with the service correctly bound to 0.0.0.0, host firewalls (ufw, firewalld, Windows Defender Firewall) or cloud security groups drop inbound traffic to ports that were never explicitly opened, which produces the same connection-refused or connection-timeout behavior from the client side.

Fix: Open port 11434 for the specific source network you trust, not for the whole internet. On Ubuntu use ufw, on a cloud VM update the security group or NSG rule.

sudo ufw allow from 10.0.0.0/8 to any port 11434
sudo firewall-cmd --add-port=11434/tcp --permanent

If a previous Ollama instance did not shut down cleanly, or another application was configured to use the same port, the new Ollama process fails to bind and can exit silently in some install methods, leaving nothing listening even though you just ran the start command.

Fix: Check what is bound to the port and stop the stale process, or change Ollama's port with OLLAMA_HOST=127.0.0.1:11435 and update your client to match.

sudo lsof -i :11434
sudo kill <pid>

A container that calls localhost:11434 is talking to its own network namespace, not the host, so even a healthy host-level Ollama service is unreachable. This produces connection refused inside the container while the host itself works fine.

Fix: Use host.docker.internal instead of localhost from inside the container on Docker Desktop, run the container with --network host on Linux, or point OLLAMA_HOST at the container-reachable host IP.

curl http://host.docker.internal:11434/api/tags
docker run --network host ollama/ollama