Deploy and Host Open WebUI + LiteLLM on Railway

Team chat over your models: Open WebUI talks only to LiteLLM. This is the ChatGPT Team replacement on c17.io/stacks/open-webui-litellm.

Provider API keys are BYO and optional. The stack boots without OPENAI_API_KEY / ANTHROPIC_API_KEY. You add keys in LiteLLM (or as optional env vars) when you want models to answer.

About Hosting Open WebUI + LiteLLM

Open WebUI is the ChatGPT-style UI. LiteLLM is an OpenAI-compatible proxy: one URL, many providers, virtual keys, spend logs. Agents (Cursor, n8n, your bots) should point at LiteLLM, not at Open WebUI.

On Railway this is three services:

Service	Source	Public	Notes
`open-webui`	`ghcr.io/open-webui/open-webui`	Yes	`PORT` native. Health `GET /health`. Volume `/app/backend/data`.
`litellm`	`docker.litellm.ai/berriai/litellm:main-stable`	Yes	So agents can hit the proxy. Health `GET /health/liveliness`.
`Postgres`	Railway plugin	No	LiteLLM models, virtual keys, spend. Open WebUI stays on SQLite on the volume unless you later set `DATABASE_URL`.

Open WebUI is pre-wired to LiteLLM over Railway private networking. First Open WebUI signup becomes admin; then disable public signup. LiteLLM admin UI is /ui on the LiteLLM domain; password is LITELLM_MASTER_KEY.

Common Use Cases

Replace ChatGPT Team seats: one UI, many models, your keys
Shared proxy for Cursor, n8n, and bots (OPENAI_API_BASE = LiteLLM)
Add a local Ollama later without changing the UI (LiteLLM model row)
Audit prompts on a box you own instead of a vendor log

Dependencies for Open WebUI + LiteLLM Hosting

Required (wired or generated — you do not type these)

Variable	Service	Source
`PORT`	both apps	Injected. Open WebUI reads `PORT`. LiteLLM start: `litellm --port $PORT --host 0.0.0.0` (or target port `4000`).
`WEBUI_SECRET_KEY`	open-webui	Generated. JWT / session secret.
`WEBUI_AUTH`	open-webui	`true`
`ENABLE_OLLAMA_API`	open-webui	`false` (no host Ollama on Railway)
`OPENAI_API_BASE_URL`	open-webui	`http://${{litellm.RAILWAY_PRIVATE_DOMAIN}}:${{litellm.PORT}}/v1`
`OPENAI_API_KEY`	open-webui	`${{litellm.LITELLM_MASTER_KEY}}` (the proxy key, not OpenAI)
`LITELLM_MASTER_KEY`	litellm	Generated. Prefix `sk-`. Admin UI password.
`LITELLM_SALT_KEY`	litellm	Generated. Encrypts stored provider keys.
`STORE_MODEL_IN_DB`	litellm	`True`
`DATABASE_URL`	litellm	`${{Postgres.DATABASE_URL}}`

Optional — BYO provider keys

None of these are required to boot. The UI comes up with an empty model list until you add a key in LiteLLM or set one of these.

Variable	Required	Why
`OPENAI_API_KEY`	No	OpenAI / compatible. Set on litellm, then add the model in LiteLLM UI (`os.environ/OPENAI_API_KEY`).
`ANTHROPIC_API_KEY`	No	Anthropic. Same pattern.
`GEMINI_API_KEY`	No	Google Gemini. Same pattern.
`AZURE_API_KEY` / `AZURE_API_BASE`	No	Azure OpenAI. Same pattern.
`OPEN_WEBUI_DATABASE_URL`	No	Only if you move Open WebUI off SQLite. Do not reuse LiteLLM's `DATABASE_URL` — different schemas.

Volume

Mount	Service	Why
`/app/backend/data`	open-webui	Users, chats, uploads, SQLite. Without it, the admin account resets every deploy.

LiteLLM state is in Postgres. It does not need a volume.

Why Deploy Open WebUI + LiteLLM on Railway?

ChatGPT Team is $30/user/month and you still cannot point it at last week's bench winner. This stack is a UI + a proxy. You pay Railway plus the tokens you burn. Ten seats of ChatGPT Team is the ~$3,600/year line on the c17 stack page. Trial credits cover the first deploy for most people. c17 never charges you — Railway usage is the referral / template kickback.

See TEMPLATE.md for the marketplace / composer shape.

Run locally

cp .env.example .env
# set WEBUI_SECRET_KEY, LITELLM_SALT_KEY (openssl rand -hex 32)
# set LITELLM_MASTER_KEY=sk-$(openssl rand -hex 32)
# set POSTGRES_PASSWORD
# optional: OPENAI_API_KEY / ANTHROPIC_API_KEY
docker compose up

Open WebUI: http://localhost:3000
LiteLLM admin: http://localhost:4000/ui (user admin, password = LITELLM_MASTER_KEY)

The UI works without provider keys. Chats fail until you add a model in LiteLLM.

License

Open WebUI and LiteLLM keep their upstream licenses. This repo is config and docs only.