A working browser page does not prove Open WebUI's backend can reach Ollama. Draw two boxes first: where Open WebUI runs, and where Ollama runs. The address must make sense from the first box.
Choose the address for your layout
- Both applications running directly on the same computer: try
http://localhost:11434. - Open WebUI in Docker, Ollama on the host: the documented host address is
http://host.docker.internal:11434. On Linux Docker, the documented setup includes--add-host=host.docker.internal:host-gateway. - Both in a container network: use the Ollama service/container hostname on that network. Do not substitute a host address without checking the actual deployment.
- Bundled Open WebUI/Ollama image: use its bundled configuration, not a connection aimed at a different host installation.
Diagnose before changing anything
Confirm Ollama answers /api/tags on its own machine. Then open Open WebUI's admin Settings → Connections and inspect the Ollama connection URL. Use the connection's verification control. Ensure you are editing an Ollama connection, not an OpenAI-compatible connection that expects another API shape.
For Docker, docker ps shows running containers and docker logs CONTAINER_NAME --tail 80 shows recent logs. Replace the name with the one from the listing. Redact secrets and private prompts before sharing output.
A common networking trap
A host service listening only on loopback may remain unreachable from a container. The official guide describes configuring Ollama's listening address, but broadening it is a security change, not a harmless universal fix. Do not open the API to the internet or disable authentication/CORS indiscriminately. For a beginner's single-computer setup, prefer the documented bundled arrangement or a native local client rather than exposing a host service.
Confirm recovery
Refresh the model list, select an installed model and send a harmless short prompt. If the connection verifies but no models appear, inspect the model inventory on the specific server you connected to.
Documentation checked November 14, 2026. Deployment-specific networking has not been hands-on tested here.