Quickstart

From zero to a working local AI in five minutes. Install the engine, pull a model, ask your first question.

Install

Clone the repo, create a Python environment, and install the engine with the extras you need.

git clone https://github.com/atypisch/soriku.git
cd soriku
python -m venv .venv && source .venv/bin/activate
pip install -e ".[all]"

Soriku is a Python 3.11+ FastAPI application. The [all] extra pulls in server, MCP, document parsing, billing and voice dependencies. If you want a leaner install pick a smaller extra like [server].

Pull a model through Ollama

Soriku routes prompts to the model that scores best for the task. At least one local model needs to be available before it can route anywhere.

ollama pull qwen2.5-coder:7b
ollama pull deepseek-r1:7b
ollama pull qwen2.5:7b
ollama pull nomic-embed-text

The first three cover code, reasoning and general use. nomic-embed-text is the default embedding model and stays loaded permanently because it's tiny.

Start the server

python server.py

The chat UI is now at http://localhost:8765. The OpenAI-compatible endpoint lives at http://localhost:8765/api/v1/chat/completions.

Ask your first question

Open the UI, type a prompt, and Soriku picks the best installed model for the task. The response shows which model answered and why it was picked.

Routing without benchmarks

Out of the box, Soriku uses bootstrap scores derived from public benchmarks. To get routing tuned to your hardware run the local benchmark suite. See Run benchmarks.

Where to go next

Installation for the full install matrix (Docker, system requirements, GPU detection).
Smart routing to understand what Soriku is doing when it routes.
Add a remote provider to bring Claude, OpenAI, Mistral, Groq or Gemini into the routing pool.
Use with Cursor or VS Code for the MCP integration.