An open-source project called Songe is trending on GitHub this week, offering a new way for people who run AI models on their own computers to make those models better at their specific tasks without sending data anywhere. The engine, developed by sous-labs, connects to local model servers like LM Studio and coding agents such as opencode and pi, then quietly trains personalized model versions while the user is away from the keyboard.
How It Plugs Into Your Workflow
Songe sits on a local proxy at port 7613, positioned between the agent and the model server. As a user works normally — correcting the model, stating preferences, or watching a tool fail — Songe captures those moments. It does not train on the raw conversation. Instead, it extracts the teachable moments: the correction, the retry, the error. When the agent goes idle, memory is free, and battery life is sufficient, Songe begins a training cycle the project calls a Dream. The process pauses instantly the moment the user returns.
Two Kinds of Dreams
The system uses two tiers of learning. A Quick Dream takes only minutes; it reads the captured lessons and injects them into the model’s system instructions, adjusting how the model behaves without changing its weights. A Deep Dream goes further. It trains a small adapter — a lightweight add-on to the model — using a technique called on-policy distillation. The original model acts as the teacher, generating practice scenes in the agent’s own tools. The adapter, the student, learns from those scenes while rehearsing earlier skills so they are not forgotten. The original model files are never modified; learned versions are stored in a dedicated songe/ folder inside LM Studio’s models directory.
Keeping Regressions Out
Before a new version replaces the one in use, a judge model compares them at the exact points where the user previously reacted — corrections, retries, errors. The new version is deployed only if it performs clearly better, does not break what already worked, and keeps its answers healthy. If anything looks off, the update is discarded. Users can also roll back to any previous version with one click.
Hardware Realities
Training requires a GPU. Machines without one can still serve models and record conversations for later learning, but they cannot run the training step itself. The project’s documentation is explicit about this constraint, and the team maintains a community testing board where users share results from different hardware configurations.
Privacy by Architecture
There is no account creation, no telemetry, and no cloud connectivity. All data — conversations, lessons, trained adapters — stays in a local folder: ~/Library/Application Support/songe on macOS, $XDG_DATA_HOME/songe on Linux, or %LOCALAPPDATA%\songe on Windows. Songe never executes the agent’s tools; it only observes the traffic between agent and model.
Installation and Daily Use
Users can install Songe via uv tool install songe, pipx install songe, or Homebrew (which requires a one-time brew trust sous-labs/tap step). The songe init command detects the local model server and agents, then shows what it will configure. A background service starts at login and auto-updates through the package manager during quiet moments. Running songe uninstall restores the agent’s original settings; adding the --purge flag removes all learned data and versions.
What Comes Next
The project’s roadmap and testing board are open on GitHub, where the team is actively seeking community hardware reports to broaden compatibility. As of November 20, 2026, the repository is trending on GitHub.
Sources
Photo: BalticServers.com (CC BY-SA 3.0)




