GoModel & Open WebUI

Overview

Open WebUI is a self-hosted chat UI, but on its own it cannot tell you what each user’s chats cost or keep an audit trail of model traffic. Pointing its OpenAI connection at GoModel fixes that: one connection, and every chat is logged, priced, and attributed to the signed-in Open WebUI user. The same connection puts Anthropic, Gemini, and every other provider configured in GoModel into Open WebUI’s model picker. Flow: Open WebUI chat -> Open WebUI backend -> GoModel -> OpenAI/Anthropic/Gemini/... What you get:

Full request and response bodies for every chat in the GoModel audit log.
Cost per Open WebUI user via forwarded user headers and labels.
All providers behind one OpenAI connection; no per-provider setup in Open WebUI.
Prometheus metrics and OpenTelemetry traces for chat traffic, once you enable those exporters (see Notes).
Failover, rate limits, budgets, and response caching applied to chat traffic, when you configure them in GoModel.

1. Run GoModel next to Open WebUI

Run both services on one Docker network so Open WebUI can reach GoModel by container name:

docker-compose.yml

services:
  gomodel:
    image: enterpilot/gomodel
    environment:
      GOMODEL_MASTER_KEY: change-me
      OPENAI_API_KEY: sk-...
      ANTHROPIC_API_KEY: sk-ant-...
      TAGGING_HEADER_1: X-OpenWebUI-User-Email
      TAGGING_HEADER_1_DONOTPASS: "true"
    ports:
      - "127.0.0.1:8080:8080"
    volumes:
      - gomodel_data:/app/data

  openwebui:
    image: ghcr.io/open-webui/open-webui:v0.11.3
    environment:
      OPENAI_API_BASE_URL: http://gomodel:8080/v1
      OPENAI_API_KEY: change-me
      ENABLE_OLLAMA_API: "false"
      ENABLE_FORWARD_USER_INFO_HEADERS: "true"
    ports:
      - "3000:8080"
    volumes:
      - openwebui_data:/app/backend/data

volumes:
  gomodel_data:
  openwebui_data:

Variable	Why
`OPENAI_API_BASE_URL`	Sends all OpenAI-API traffic to GoModel. Keep the `/v1` suffix.
`OPENAI_API_KEY`	A GoModel API key (or `GOMODEL_MASTER_KEY`), not a real OpenAI key.
`ENABLE_OLLAMA_API`	Off unless you run Ollama; avoids connection errors in the logs.
`ENABLE_FORWARD_USER_INFO_HEADERS`	Adds the signed-in user’s name, email, id, and role as headers on every model call, which GoModel turns into labels below.
`TAGGING_HEADER_1`	Tells GoModel to record that header’s value as a label on every request.
`TAGGING_HEADER_1_DONOTPASS`	Keeps the label but strips the header before forwarding, so user emails never reach a provider on passthrough routes. Translated routes never forward client headers.

Open http://localhost:3000, create the first (admin) account, and the model picker lists exactly what GoModel exposes — including Anthropic and Gemini models and any virtual models. Existing installations can set the same connection at runtime under Admin Panel -> Settings -> Connections -> OpenAI API instead of using environment variables.

Create a dedicated API key for Open WebUI in the GoModel dashboard (API Keys -> Create API Key) instead of using the master key. Keys carry labels, budgets, and rate limits, so chat spend shows up separately in the dashboard.

2. Chat, and see every call in GoModel

Pick any model — openai/gpt-5-mini, anthropic/claude-haiku-4-5-20251001, gemini/gemini-2.5-flash — and chat. Streaming responses, and the follow-up title and tag generation Open WebUI runs after each chat, all pass through GoModel. Each call appears in the audit log with the full prompt, response, tokens, cost, and the headers identifying the user.

Verified with Open WebUI 0.11.3: model list, streaming and non-streaming chat, title generation, and per-user labels, against OpenAI, Anthropic, and Gemini models routed through GoModel.

Per-user cost tracking

With the two variables from step 1 set, every request carries X-OpenWebUI-User-Email (plus -Name, -Id, and -Role) and GoModel records the email as a label. The dashboard’s usage-by-label view then shows requests, tokens, and cost per person. Add TAGGING_HEADER_2: X-OpenWebUI-User-Name (with its own TAGGING_HEADER_2_DONOTPASS: "true") for friendlier labels. If emails should not appear in usage data, set TAGGING_HEADER_1: X-OpenWebUI-User-Id in step 1 instead, keeping TAGGING_HEADER_1_DONOTPASS. Header-name matching is case-insensitive, so the X-Openwebui-User-Email casing Open WebUI actually sends is fine. Keep TAGGING_HEADER_1_DONOTPASS set: regular routed requests never forward client headers, but passthrough requests would otherwise carry the email header to the provider. See Labelling for prefixes and keeping these headers from reaching the upstream provider.

Tame the model picker

GoModel can expose hundreds of models, and Open WebUI lists all of them. To keep the picker usable:

Hide models or restrict them to user groups under Admin Panel -> Settings -> Models.
Set DEFAULT_MODELS so new users start on a model you chose.
Or expose fewer providers in GoModel; the picker follows GET /v1/models.

Pick a cheap task model

After each chat, Open WebUI generates a title and tags using the same model as the chat by default — a few hundred extra prompt tokens per conversation, which you will now see itemised in the audit log. Set a small model (for example openai/gpt-5-mini) as the task model under Admin Panel -> Settings -> Interface to keep that overhead cheap. Task calls go through GoModel too, so they stay logged and priced.

Embeddings and RAG

Open WebUI’s document RAG defaults to a local embedding model. To route embeddings through GoModel instead, set the embedding engine to openai under Admin Panel -> Settings -> Documents with the same base URL and key as step 1. GoModel serves /v1/embeddings and prices embedding usage like any other call. Switching engines requires reindexing existing documents; chat routing works fine without this.

Troubleshooting

Symptom	Fix
Connection error when verifying the OpenAI API connection	Open WebUI cannot resolve the base URL. Use the Docker service name (`gomodel`), not `localhost`; `localhost` inside the container is the container itself.
`401 Unauthorized` in GoModel logs	The connection key does not match a GoModel API key or `GOMODEL_MASTER_KEY`.
Model picker is empty	GoModel has no provider configured. Check `GET /v1/models` and provider environment variables, then refresh the connection under `Admin Panel -> Settings -> Connections`.
No per-user labels in usage	`ENABLE_FORWARD_USER_INFO_HEADERS` is off, or `TAGGING_HEADER_1` is not set in GoModel. Both are needed.
Ollama connection errors at startup	No Ollama is running. Set `ENABLE_OLLAMA_API: "false"`.
Same answer repeats for identical prompts	GoModel’s response cache replays identical requests. This is usually what you want; disable the cache if not.

Notes

GoModel failover and budgets apply to chat traffic automatically; a per-key budget caps what the whole Open WebUI instance can spend.
Image generation, speech-to-text, and text-to-speech have separate endpoints in Open WebUI’s admin settings and were not part of the verification run.
For dashboards and alerts, see Prometheus metrics and OpenTelemetry.

​Overview

​1. Run GoModel next to Open WebUI

​2. Chat, and see every call in GoModel

​Per-user cost tracking

​Tame the model picker

​Pick a cheap task model

​Embeddings and RAG

​Troubleshooting

​Notes

Overview

1. Run GoModel next to Open WebUI

2. Chat, and see every call in GoModel

Per-user cost tracking

Tame the model picker

Pick a cheap task model

Embeddings and RAG

Troubleshooting

Notes