Problem: Open WebUI in Docker cannot connect to Ollama (Ollama not connected / empty model list) —— Trove problem library

What you will run into

After deploying Open WebUI with Docker, the page keeps showing "Ollama not connected" and the model list is empty; inside the container, localhost:11434 returns Connection refused.

The most common cause: following a "local install" tutorial and setting OLLAMA_BASE_URL to http://localhost:11434 — wrong inside a container.

Why this happens

Open WebUI runs inside a Docker container, where localhost is the container itself, not the host.

Ollama listens on the host's 127.0.0.1:11434 by default; nothing listens on localhost:11434 inside the container → connection refused (measured: Errno 111 Connection refused).

To reach the host's Ollama, use the host address as seen from the container network:

Note: newer Open WebUI uses the plural env var OLLAMA_BASE_URLS (older docs use singular OLLAMA_BASE_URL); wrong variable name also fails to connect.

Open WebUI runs in a container and Ollama runs on the host — the network-namespace rule "localhost inside a container is the container itself" (see docker-host-access) shows up here directly as Connection refused.

The config layer has a version-evolution trap too: newer Open WebUI uses the plural key OLLAMA_BASE_URLS (older versions used singular OLLAMA_BASE_URL) — a wrong key name fails silently, and the UI still reports "cannot connect".

Deterministic takeaway: on connection failure, first separate "addressing layer" from "config-key layer" — if curl http://host.docker.internal:11434 works inside the container, addressing is fine and the problem is in Open WebUI's key/value (URL with http:// prefix, plural key name);

Docker Desktop injects host.docker.internal by default while Linux engines need --add-host host.docker.internal:host-gateway; after changing env vars, restart the container (env is only read at startup).

Step by step

1. Confirm it is the "in-container address" issue (30-second check) — Required (confirm before touching anything)

docker exec open-webui python3 -c "import urllib.request
for u in ['http://localhost:11434/api/version', 'http://host.docker.internal:11434/api/version']:
    try: print(u, '-> OK:', urllib.request.urlopen(u, timeout=5).read.decode.strip)
    except Exception as e: print(u, '-> FAIL:', str(e)[:60])"

Measured output shape: localhost:11434 -> FAIL: Connection refused, host.docker.internal:11434 -> OK: {"version":"0.32.14"}. Match = this pit.

2. Check the container start arguments — Diagnose (missing --add-host / wrong var name?)

docker inspect open-webui — look at Config.Env for OLLAMA_BASE_URLS pointing to localhost, and HostConfig.ExtraHosts for a host-gateway mapping (Linux engine).

3. Fix per the table and recreate — Required (recreate applies the fix)

• Value: http://host.docker.internal:11434 (no trailing slash);
• Variable: use OLLAMA_BASE_URLS (plural) on new versions;
• Linux: add --add-host host.docker.internal:host-gateway; compose uses extra_hosts.

4. Verify — Required (check from inside and outside the container)

• Container view: host.docker.internal:11434 returns the Ollama version (measured);
• Page: after refresh, the model dropdown lists the host Ollama models (tool-side integration measured: model discovery passed).

How to confirm it works

docker exec open-webui python3 -c "import urllib.request; print(urllib.request.urlopen('http://host.docker.internal:11434/api/tags', timeout=5).read[:200])" returns the model list;

the page model dropdown shows host models (e.g., qwen2.5-0.5b, nomic-embed-text).

Mirror case: external devices / other containers can't reach Ollama

The mirrored trap: the address in the UI is correct, but the Ollama server itself only listens on 127.0.0.1 — a phone on the LAN, another machine, another container, or a remote IDE all fail with connection refused.

Ollama binds loopback by default; to expose it you must change the listen address:

# Server side: listen on all interfaces (LAN / container networks reachable)
OLLAMA_HOST=0.0.0.0:11434 ollama serve
# systemd install: add to [Service] in /etc/systemd/system/ollama.service
#   Environment="OLLAMA_HOST=0.0.0.0:11434"
# Docker install: add -e OLLAMA_HOST=0.0.0.0:11434

Verify from outside: curl http://<host-ip>:11434/api/version returns {"version":"..."}.

Two caveats: 0.0.0.0 means the API is open to the LAN (ollama has no auth by default — anyone can list models / run inference; don't expose it in production);

older releases had a regression that bound localhost only (ollama/ollama#12053, fixed in v0.11.5/0.11.6) — upgrading is also a fix.

Other environments

Symptom

Docker-deployed Open WebUI shows "Ollama not connected" with an empty model list; accessing localhost:11434 from inside the container is refused.

Cause

localhost inside a container is the container itself, not the host; OLLAMA_BASE_URLS pointing to localhost:11434 cannot reach the host's Ollama. Use host.docker.internal instead (Linux engines need a host-gateway mapping).

Solution

1. Measure both addresses from inside the container: localhost refused, host.docker.internal reachable;
2. Recreate the container: -e OLLAMA_BASE_URLS=http://host.docker.internal:11434 (plural var on new versions);
3. Linux: add --add-host host.docker.internal:host-gateway (compose: extra_hosts);
4. Refresh the page and confirm the model list appears.

Related canonicals

Read in order: precondition -> same root / same domain / related:

• Related (similar errors, different cause):
• Open WebUI RAG retrieval not working / embedding misconfig (RAG_OLLAMA_BASE_URL is a separate key: chat connected != retrieval connected)
• Forgot the admin password / can't log in (WebUI account/login layer)
• ollama reports "could not connect" (first confirm the ollama server itself is running)