Skip to main content

Open WebUI

Public API

Open WebUI is a self-hosted chat front end that speaks the OpenAI protocol to whatever server you give it. Two settings, the base URL and the key, point it at your tenant's public API. Its model picker then fills with the models your tenant serves, and every conversation runs through your tenant's AI Gateway with its tiers, guardrails, budgets and logs.

Who can use it: anyone holding an API key with the Chat and Read models scopes; Open WebUI lists the models with the key before the first chat. The tenant must be licensed for the public API.

Prerequisites​

  • Docker and the ghcr.io/open-webui/open-webui:main image. Verified with Open WebUI 0.11.3.
  • Your tenant's public API address, https://<your tenant host>/public/v1, from the top of Studio's Developer Hub.
  • A key from the Developer Hub's API keys tab with the Chat and Read models scopes.

Connect​

docker run -d --name open-webui -p 3000:8080 \
-e OPENAI_API_BASE_URL=https://<your tenant host>/public/v1 \
-e OPENAI_API_KEY=kapi_YOUR_KEY \
-e ENABLE_OPENAI_API=True \
-e ENABLE_OLLAMA_API=False \
-e WEBUI_AUTH=False \
ghcr.io/open-webui/open-webui:main

WEBUI_AUTH=False skips account creation for a single-user trial on your own machine. Leave it out for anything shared, and Open WebUI asks for an account on the first visit. ENABLE_OLLAMA_API=False stops it looking for a local Ollama it will not find. Open WebUI copies the two connection values into its own database on first start, so change them afterwards in its administrator settings under connections, not in the environment.

Open http://localhost:3000. The first start takes a few minutes while Open WebUI downloads the local embedding model for its document features. Start a new chat: the model picker lists your tenant's models, with the first one selected. Choose a chat model, send a message and read the reply.

Verify​

  1. The model picker lists the same ids as your tenant's models route, plus Open WebUI's own arena entry; nothing from any other provider appears.
  2. A message gets an answer. Open WebUI streams replies by default; a request through its own chat completions API with streaming off is answered too.
  3. The key's row on the Developer Hub's API keys tab shows a new last-used time, and each turn appears in the Control Room's AI Gateway under Logs as a line naming the provider, model, tier, tokens, cost and latency. Open WebUI also asks the model for a chat title and tags after a turn when those features are on in its interface settings, so the log can show more lines than messages.

What does not carry over​

  • Agents, sessions and knowledge. Open WebUI speaks only to the model routes. Your tenant's agents, sessions and knowledge are not reachable from it, and a document uploaded into Open WebUI is embedded and stored by Open WebUI itself, not by your tenant. To chat with your agents from a front end, use the public API's agent routes or the MCP server.
  • Who made the call. Every person using the instance shares the one key, so your tenant records all of their usage against that key and its identity. Attribution per person is not available through this path.
  • Streaming pace. The reply arrives in one burst after the guardrails have read it, not word by word.
  • Which models appear. The picker shows every model the key can list, chat models by default. Hide the ones you do not want offered in Open WebUI's own model settings; the tenant does not filter the list per key.

Troubleshooting​

  • The model picker is empty. The key lacks the Read models scope, or the base URL does not end in /public/v1. Fix the value in the administrator settings under connections and reload.
  • Server Connection Error when a message is sent. The key lacks the Chat scope, or the address returned something Open WebUI could not parse. Check the base URL; curl https://<your tenant host>/public/v1/models with the key in an Authorization header should return JSON.
  • Changing the environment variables has no effect. Open WebUI copied the first values into its database. Change them in the administrator settings under connections, or remove the container's data and start again.
  • The container stays unhealthy for minutes after starting. The first start downloads a local embedding model. Wait for it, or mount a volume at /app/backend/data so the download survives a restart.

Describes the platform as of .

Was this page helpful?