Project · AI tooling

The Wizard's Familiar

A coding agent that runs entirely on my machine. One standard-library Python file, no dependencies, and a model served locally by llama.cpp.

Latest release ↗ Source ↗

Role

Solo

Stack

Python 3.11+ (stdlib only) · local model server

Status

Released publicly; Windows, Linux and macOS core

License

MIT

Links

GitHub

100%local, on-device

0runtime dependencies

3supported core platforms

$0per token

Why I built it

My team paused cloud AI while we reworked cost limits. I built a coding agent that runs on my own laptop, so I could keep working without a hosted model or a per-token bill. I still use it.

What it is

familiar.py is the whole agent: standard-library Python, streaming tool calls, file and search tools, shell execution, and context management. Its core supports Windows, Linux and macOS with Python 3.11+. You start a compatible loopback model server separately. The optional serve.sh launcher still targets macOS on Apple Silicon with a separately provisioned Ornith-1.0-35B model.

Install the wheel from GitHub Releases, run the Python file from the source archive, or use python -m pip install . from a clone. No model is downloaded by the client, and the Python runtime dependency list stays empty. There is no PyPI release.

familiar.py · CLI→ local llama-server→ your local model→ reads & edits your files

No hosted inference or API key. Normal mode executes commands as you without a sandbox or approval prompt. Plan mode dispatches only native read, list and search tools.

What I like about it

It is small on purpose. I can read the client end to end, which matters for a tool that edits my files. Plan mode enforces a read-only dispatch boundary; manual compaction summarizes older observations in memory, and context pauses before it would lose messages. Conversations are not saved, while optional input history has its own setting.

← All projects Get in touch →