Open WebUI

Rendered from docs/clients/open-webui.md at commit b9d69f7299ed. The repository file is authoritative.

Verified against the Open WebUI docs and main source on 2026-10-11 (sources in research/clients-2026-10-11.md, §1). Test status: configuration keys documented; a live Open WebUI tool round trip through ModelSaar is NOT RUN in this build.

Generate the current snippet: modelsaar client-config open-webui [--docker].

Protocol and URL

Open WebUI sends Chat Completions and string-concatenates paths onto the connection URL ({url}/chat/completions, {url}/models). Therefore the URL must include /v1 and must not end with a slash: http://127.0.0.1:8484/v1. Models are auto-listed from GET /v1/models, which ModelSaar serves truthfully (aliases only when backed by a usable deployment).

Steps (UI)

Admin Panel → Settings → Connections → Manage OpenAI API Connections → Add.
URL http://127.0.0.1:8484/v1 (from Docker: http://host.docker.internal:8484/v1).
API Key: contents of ~/.modelsaar/keys/api_keys (first line).
Save → Verify Connection (this is GET /models with the Bearer key).
Pick auto (or auto-code, auto-vision, a concrete ollama/... id) in the chat model selector.

Environment variables (Docker / compose)

ENABLE_OPENAI_API=true
OPENAI_API_BASE_URLS=http://host.docker.internal:8484/v1
OPENAI_API_KEYS=<modelsaar api key>
TASK_MODEL_EXTERNAL=auto-fast
ENABLE_FORWARD_USER_INFO_HEADERS=true

TASK_MODEL_EXTERNAL sends Open WebUI's background tasks (title, tags, follow-ups, query generation — non-streaming) to auto-fast instead of the current chat model.
ENABLE_FORWARD_USER_INFO_HEADERS=true makes Open WebUI add X-OpenWebUI-Chat-Id (plus user headers). ModelSaar reads it by default (affinity.session_headers) to keep a conversation on one deployment. Alternatively add a per-connection custom header X-ModelSaar-Session: {{CHAT_ID}} without forwarding user info. Without either, chats are request-routed (stateless).

Container networking

Open WebUI in Docker cannot reach the host's 127.0.0.1. Use host.docker.internal; on Linux add --add-host=host.docker.internal:host-gateway (compose: extra_hosts). ModelSaar must then listen beyond loopback: server.host: 0.0.0.0 + security.allow_lan: true — do this only on a trusted machine/network, or run ModelSaar in the same compose network (docker-compose.yml in this repo). Podman: host.containers.internal.

Tools, images, context

Native function calling (default) sends tools; Open WebUI does not send tool_choice or parallel_tool_calls. ModelSaar routes tool requests only to tool-verified deployments.
Images are sent as base64 data URLs → routed to vision-verified deployments (auto-vision).
Per-model context/capability toggles in Admin → Models are UI-side; ModelSaar enforces the real limits server-side and returns a 422 with blockers when nothing fits.

Verify

modelsaar test                       # gateway alone
# in Open WebUI: new chat with model `auto`, ask something; then
modelsaar decisions                  # shows the request and the chosen deployment