> ## Documentation Index
> Fetch the complete documentation index at: https://docs.prismml.com/llms.txt
> Use this file to discover all available pages before exploring further.

# open-webui

[Open WebUI](https://github.com/open-webui/open-webui) gives you a full agentic demo on top of local [Bonsai 27B](/models/bonsai-27b): chat with images, tool calling against live tools, and a database with a hidden story to investigate. One command configures everything — tools, model settings, and a demo database — with nothing to click through by hand.

## Setup (once)

```bash theme={null}
source .venv/bin/activate
uv pip install open-webui
```

## Run

```bash theme={null}
./scripts/start_openwebui.sh
```

This starts `llama-server` if it isn't already running, seeds the demo (tools, model settings, demo database), and opens `http://localhost:9090`. First boot takes about a minute for database migrations; Ctrl+C stops everything it started.

Prefer the MLX backend on a Mac?

```bash theme={null}
BONSAI_BACKEND=mlx ./scripts/start_openwebui.sh
```

Only one backend runs at a time — two resident 27B models is too heavy for most machines. The MLX backend is noticeably slower per token and has no cross-request prompt cache (each follow-up reprocesses the whole conversation, image tokens included), so multi-turn chats are slower than on llama.cpp. For interactive multi-turn use, stick with the default llama.cpp backend.

## What to try

* **Vision** — click the `+` in the message box, upload a photo or screenshot, and ask about it. Follow-up questions about the same image are near-instant on the llama.cpp backend (prompt cache).
* **Weather tool** — *"What's the weather in Lisbon right now?"*
* **Web fetch tool** — *"Fetch [https://prismml.com](https://prismml.com) and summarize it."*
* **MCP servers** — Hugging Face Hub and DeepWiki are connected but off by default. Enable one from the tool menu, then try *"Find trending vision-language models on Hugging Face."*
* **The agentic analyst** — the demo ships a SQLite database of a fictional B2B company (2024–2026) with a hidden story in it. Ask: *"Our EMEA numbers looked bad in the second half of 2025. Figure out what happened and whether it was preventable."* and watch it explore the schema, run queries, and piece the answer together.
* **Thinking** — answers show a collapsible thought block. Turn thinking off for snappier replies with `BONSAI_THINKING=0 ./scripts/start_openwebui.sh`, or cap it per request.

## How it works

* The three built-in tools live in `scripts/openwebui/` as plain Python Open WebUI "Tools" and are re-seeded on every start — edit them and restart to change the demo.
* The demo database is generated into `.openwebui/demo.db` on first start.
* Chats persist in `.openwebui/` between runs; delete that directory for a factory-fresh demo.
* Configuration comes from `start_openwebui.sh` env vars on every launch (auth disabled, single backend, background title/tag/follow-up generation off so the UI doesn't keep the model busy after each reply).

## Adding your own tools or MCP servers

Edit `TOOL_SERVER_CONNECTIONS` in `scripts/start_openwebui.sh` and restart — that script, not the admin panel, is the source of truth (the demo runs with persistent config disabled, so panel edits don't survive a restart). Full guide, including the optional Brave Search MCP server: [`TOOLS.md`](https://github.com/PrismML-Eng/Bonsai-demo/blob/main/TOOLS.md) in the demo repo.

## Related

<CardGroup cols={2}>
  <Card title="Run the server" icon="server" href="/run/server">
    The raw OpenAI-compatible API underneath Open WebUI.
  </Card>

  <Card title="Bonsai 27B" icon="layer-group" href="/models/bonsai-27b">
    The model this demo is built around: vision, tool calling, thinking.
  </Card>
</CardGroup>