Getting started

This page walks through installing OCC, launching the Node for the first time, and confirming everything works. Allow about ten minutes the first time — the model download is the slow part.

Prerequisites

You need three things on your machine before installing OCC:

• Python 3.11 or later. Download from python.org. On Windows, check "Add Python to PATH" during installation.
• Git. Download from git-scm.com. All defaults are fine.
• Ollama. The OCC setup script checks for it and prompts you to install it if it's missing. You can also install it ahead of time from ollama.com/download.

Installation

Clone the repository and run the setup script for your platform:

git clone https://github.com/VickyWenSZ/occ.git
cd occ

Windows:

start.bat

macOS / Linux:

bash start.sh

The setup script is a one-time process. It walks you through the installation step by step:

• Verifies Python, Git, and Ollama are installed.
• Installs the Python dependencies for the Node.
• Detects the available VRAM on your machine and selects the appropriate Qwen model for your hardware.
• Downloads that model through Ollama.
• Generates icons and creates an OCC Node shortcut on your desktop.
• Launches the Node and opens it in your browser.

If anything is missing, the script tells you what to install and where to find it. After the first successful run, you do not need to use start again.

First launch — what gets downloaded

The model download is the longest step in the setup. The script picks the best Qwen variant for your VRAM:

Tier	VRAM	Model	Quant	Approx size
Micro	CPU only	qwen3.5:2b	Q4_K_M	~1.5 GB
Small	4 GB	qwen3.5:4b	Q4_K_M	3.4 GB
Mid	8 GB	qwen3.5:9b	Q4_K_M	5.97 GB
Large	16 GB	qwen3.5:9b	Q8_0	9.53 GB
XL	24 GB	qwen3.5:9b	BF16	19 GB
Server S	32 GB	qwen3.6:27b	Q4_K_M	16.8 GB
Server L	80 GB	qwen3.6:27b	BF16	56 GB

The download happens once. Subsequent launches start in seconds — Ollama keeps the model cached on disk.

All tiers run the same code paths. The differences are in the size of the model, the context window, and the retrieval budget — see The Node for how OCC adapts to each tier.

Daily use

After the first setup, use the OCC Node icon on your desktop. Double-clicking it:

1. Starts the local Ollama service if it isn't running.
2. Starts the Node server on localhost:7891.
3. Opens the browser tab automatically.

When you close the browser tab, the Node stops cleanly.

Manual fallback — if the desktop shortcut doesn't work for any reason (moved folder, OS quirk, custom setup):

python node/apps/gui/server.py

Then open http://localhost:7891 in your browser.

First query

When the Node opens, the chat panel is empty and waiting. Try a simple test:

Ask: What is the Inca road system?

Watch the routing label that appears next to the answer. For a knowledge question, it should show something like server + local (retrieval against the broker, deliberation locally) or distributed (retrieval against the broker, Critic step running on a peer with stronger hardware).

If you only see an Expert draft with no Critic review, the Node ran in chat mode rather than deliberate — try a more substantive factual question.

Updating

Inside OCC Node, open Settings and click Update to latest. The Node pulls the newest version of the code, reinstalls dependencies, and restarts automatically.

Manual fallback:

git pull
python -m pip install -r node/requirements.txt

Then restart the Node from the desktop shortcut.

When something doesn't work

• The browser opens but the page doesn't load. Wait ten to fifteen seconds — the server takes a moment to initialize on slow machines. If it still doesn't load, check that no other process is using port 7891.
• The model didn't download. Make sure Ollama is running (ollama list in a terminal should respond). If not, start it with ollama serve. Re-run the setup script — it will resume the download.
• An answer says "Server unavailable". The Node couldn't reach the broker. Check your internet connection. If you want to use only local packs anyway, toggle Settings → Network → Use local packs only.
• A query takes a long time on first run after restart. The model needs to load into VRAM. The first query of a session is slow (several seconds); subsequent queries are fast as long as the Node stays open.

For anything else, the Broker agent panel inside the Node UI streams the connection status, peer events, and routing decisions in real time.

What to do next

Once your Node is running:

• Read How OCC works if you want to understand what happens between your question and the answer.
• Try Forge if you want to build your own pack from a set of curated sources.
• Browse The Node for the full reference of modes, slash commands, and settings.
• Check Security and privacy for what data goes where.

Welcome to the network.