sous-labs/songe
Machine LearningAn on-device training engine that learns from your coding agent sessions and fine-tunes a local model during idle time.
What it does
Songe sits between a coding agent and a locally served model, recording what happens after each model action: tool results, errors, and your follow-up messages. When the agent goes quiet, it turns those moments into training material so the local model gets better at anticipating them. It targets people running local models through OpenAI-compatible agents such as opencode or pi.
The learning has two forms. A Quick Dream appends your stated or demonstrated preferences to the model's instructions and takes minutes. A Deep Dream writes practice scenes from the recorded lessons and trains a small adapter on the model itself, using the original model as teacher and the adapter-equipped model as student (on-policy distillation). New versions are adopted only if they clearly beat the current one on the moments you reacted to, and earlier versions remain reversible.
Quick start
You need Python 3.11+ and either uv or pipx, a locally served model, and an agent. Install as a tool:
uv tool install songe # or: pipx install songe
songe init
On macOS via Homebrew the equivalent is brew tap sous-labs/tap, brew trust sous-labs/tap, brew install songe, then songe init.
songe init discovers your model servers and agents, prints what it will change (data folder, a background service, provider entries in opencode and pi), and asks once before applying. After that, select a songe/… model in your agent and keep working; learning starts when the agent is idle. The dashboard is available through songe open, and songe sleep starts a Deep Dream immediately.
Constraints and side effects
The model server can be LM Studio, vLLM, Ollama, llama.cpp, or mlx-lm; end-to-end testing is reported for LM Studio on macOS and vLLM on Linux. Training needs Apple Silicon (MLX) or NVIDIA/AMD (PyTorch); machines without a GPU can still serve and record but will not train. pip install songe only works inside a virtual environment.
Installation writes a data folder under ~/Library/Application Support/songe, $XDG_DATA_HOME/songe, or %LOCALAPPDATA%\songe, creates a songe/ folder inside LM Studio's models folder, adds one provider entry to opencode and pi with a backup, and registers a background service that starts at login. songe forget deletes recorded conversations while keeping learned versions; songe uninstall stops the service and restores agent settings, with --purge additionally removing data and learned versions. The project is Apache-2.0 licensed.
HotRepo's take
The repository is new: created on 2026-11-20, it recorded about 107 stars and 15 forks by the next day, with a new-list snapshot showing 7 stars added over a one-hour window on 2026-11-21. That is too short a series to say anything about sustained traction. The interesting technical detail is the scope: the engine is designed to keep data local, avoid training on raw conversations directly, and gate new adapters behind comparison checks, which addresses the main practical concern of running fine-tuning next to daily work on a single machine.
License
Apache-2.0.
Written by AI from the project's public information and HotRepo's records, updated . The project's own docs are the source of truth.