Generate code with a local model

Describe a program in plain language, and a model on your own computer writes it, runs it and checks it. There is no cloud service and no API key. A coding agent, pi or Open Interpreter, talks to a model that Ollama serves on your machine. The satyrn command downloads them in pinned, checksum-verified versions into a single folder, helps you choose a model that fits your computer, and starts and stops the model server for you.

Quick start

You need git and Python 3.14 (or uv, which fetches Python for you). Clone the project and install the satyrn command:

Linux and macOS

git clone https://github.com/niccokunzmann/satyrn-getting-started.git
cd satyrn-getting-started
python3.14 satyrn.py install   # the `satyrn` command and its shell completion
satyrn                         # shows what is missing and guides you through it

Windows (PowerShell)

git clone https://github.com/niccokunzmann/satyrn-getting-started.git
cd satyrn-getting-started
py -3.14 satyrn.py install     # the `satyrn` command and its shell completion
satyrn                         # shows what is missing and guides you through it

uv

git clone https://github.com/niccokunzmann/satyrn-getting-started.git
cd satyrn-getting-started
uv run --no-project --python 3.14 satyrn.py install   # uv fetches Python 3.14 for you
satyrn                         # shows what is missing and guides you through it

satyrn checks this computer, lists the models that fit it, and offers each missing step with a default answer: download the tools, download a model, start the model server. Press Enter to accept the defaults. Then start working in a folder of your choice:

satyrn code ~/ai-code

Type what you want, for example:

Create hello.py that prints Hello, World! Then run it.

The model writes the file into ~/ai-code, runs it and tells you what it did. The animation shows the important moments of a real session with Open Interpreter; see the whole session as text. When you are done, leave with Ctrl+C and stop the server with satyrn stop. Nothing starts by itself, and everything the tools download stays in the satyrn-getting-started folder.

How fast is it?

On a laptop CPU without a GPU (Intel i7-13700H), the default model gpt-oss:20b writes and checks a small program like the one above in about two minutes. A GPU is used automatically if Ollama finds one. Choose a model compares four models.

Use it

• Set up
  • What you need
  • Python 3.14
  • Install the satyrn command
  • Let satyrn guide you
  • Choose the coding agent
  • The same steps by hand
  • Defaults
  • Check that it works
  • Stop and remove
  • Troubleshooting
• Generate code
  • Talk to it
  • Example: hello world
  • One request, no conversation
  • Write good requests
  • Use the coding agents directly
  • Stop
• Choose a model
  • Which models fit your computer
  • Select a model
  • Benchmark
  • Reading the context before your request
  • Benchmark your own computer
  • Add a model that is not listed

Background

• How it works
  • The pieces
  • Two coding agents
  • What the model is told
  • Runs that cannot finish
  • Where everything lives
• Speed: how to get the first answer sooner
  • Result
  • Overview
  • 1. Loading the model
  • 4. Ollama settings: batch size and threads
  • 2. Reading the context in advance (warm-up)
  • 3. The size of the context per coding agent
  • 6. Open Interpreter’s extra title request
  • 7. Less reasoning
  • 5. Flash attention and a compressed context cache
• Reproducibility
  • When a model has changed
  • Updating a pin
• Satyrn’s own engine (Apple Silicon)
  • Setup
• The Satyrn projects
• Reference
  • satyrn commands
  • make targets
  • Files you may edit
  • Environment
• Contributing
  • Git hooks
  • The documentation
  • Goals and open work