Open WebUI with an OpenAI-compatible endpoint

Open WebUI is a self-hosted chat interface that teams put in front of models. It connects to OpenAI-compatible endpoints through environment variables, supports several endpoints at once, and — this is the part that catches everyone — stores those values as persistent config, so changing the environment later does nothing. This page covers all three, plus what a shared deployment costs per person.

One endpoint: two environment variables

For a single endpoint set OPENAI_API_BASE_URL and OPENAI_API_KEY. The base URL ends in /v1; Open WebUI appends the rest of the path itself.

# docker-compose.yaml — one endpoint, values read from an .env file
# sitting next to this one (and kept out of git).
services:
  open-webui:
    image: ghcr.io/open-webui/open-webui:main
    ports:
      - "3000:8080"
    environment:
      - 'OPENAI_API_BASE_URL=${GATEWAY_BASE_URL}'   # https://aicomp.ai/v1
      - 'OPENAI_API_KEY=${GATEWAY_API_KEY}'
    volumes:
      - open-webui:/app/backend/data
volumes:
  open-webui:

Keeping the values in an .env file beside the compose file rather than inline means the compose file is safe to commit and the key is not. That matters more here than in a single-user tool, because these deployments usually end up in a shared repository.

Several endpoints: the lists pair by position

The plural variables, OPENAI_API_BASE_URLS and OPENAI_API_KEYS, are both semicolon-separated lists. They are matched by position — the first URL is used with the first key, the second with the second. There is no key-to-URL mapping beyond order, so a missing separator silently pairs every subsequent entry with the wrong key.

# Multiple endpoints: semicolon-separated, paired BY POSITION.
# The first URL uses the first key, the second URL the second key.
export OPENAI_API_BASE_URLS="https://aicomp.ai/v1;http://localhost:11434/v1"
export OPENAI_API_KEYS="sk-your-key;unused-for-local"

Local models typically ignore the key, but the slot still has to exist. A placeholder is fine; an empty slot is not, because it shifts the alignment for everything after it.

The trap: these values are persistent

OPENAI_API_BASE_URL and OPENAI_API_KEY are persistent config values. The first time Open WebUI starts, it writes them into its database. After that, changing the environment variable has no effect — the stored value wins. People restart containers for an hour wondering why the new endpoint is not being used.

To change an endpoint after the first run, change it in the admin connections UI, or reset the stored value. The environment variables are authoritative once, at first start.

Model list comes from the endpoint

The model dropdown is populated from /v1/models. An empty dropdown is a connection problem, not a missing-models problem — worth checking directly instead of reloading the interface.

# The model dropdown is populated from /v1/models. If it is empty,
# ask the endpoint directly rather than reloading the UI.
curl "$GATEWAY_BASE_URL/models" \
  -H "Authorization: Bearer $GATEWAY_API_KEY"

# Empty or 401 here means the UI can never populate either.

What a shared deployment costs

Self-hosting spreads one endpoint across a team, so the useful number is cost per person rather than cost per request. The table below assumes a five-person team at thirty chats each per working day, nine thousand input tokens and fourteen hundred output tokens per chat.

How this fails in practice

FAQ

Related

• Every tool, one endpoint — the full matrix
• Put a proxy in front if you want routing and budgets
• The same endpoint inside the IDE
• Track what the team actually spends
• What 'OpenAI-compatible' actually guarantees
• All model rates

Get API access