This release is a pre-release and may not be stable for production use.
A training engine for your local model, using your idle compute.
Install • How it works • Commands • Roadmap • Website
Use your local model as usual through opencode, pi or any OpenAI-compatible tool. While you work, Songe notices what happens after each of the model's actions: the tool result, the error, your next message. When your agent is quiet, it teaches the model to see those coming. A new version is used only if it passes its checks, and nothing ever leaves your machine.
✨ Highlights
- Learns from how you work. Your corrections, stated preferences and the tools' errors become lessons for your own model.
- Two kinds of Dreams. A Quick Dream adds your preferences to the model's instructions in minutes; a Deep Dream trains the model itself while you are away.
- Safe by default. A new version is used only if it does better on your own conversations; one click goes back to any earlier version.
- Stays out of the way. Runs only when your agent is quiet, memory is free and the battery is fine, and stops the moment you are back.
- Private. No account, no telemetry, no cloud. Songe never runs your tools.
📥 Install
With uv or pipx (Python 3.11 or later):
uv tool install songe # or: pipx install songe
songe init
macOS with Homebrew:
brew tap sous-labs/tap
brew trust sous-labs/tap # newer Homebrew asks this once
brew install songe
songe init
songe init finds your model server and your agents, shows what it will
change, and asks once:
$ songe init
Models 3 from LM Studio, each as songe/…
Agent opencode → Songe models · default songe/granite-4.2-3b-MLX-4bit
Agent pi → Songe models · choose one with /model
Data ~/Library/Application Support/songe
Service starts at login
Learning new conversations · starts when idle
Replace loaded copy when a better version is ready
Install? (Y/n)
Then pick a songe/… model in your agent, and work as usual.
Plain pip, and what you need
pip install songe works inside a virtual environment; outside one,
Homebrew's and most Linux distributions' Python refuse it.
You need a model served locally, and an agent:
| Supported | Tested end to end | |
|---|---|---|
| Model servers | LM Studio, vLLM, Ollama, llama.cpp, mlx-lm | LM Studio (macOS), vLLM (Linux) |
| Agents | opencode and pi (set up for you); any OpenAI-compatible tool | opencode, pi |
| Training | Apple Silicon (MLX); NVIDIA and AMD (PyTorch) | Apple Silicon, NVIDIA |
Machines without a GPU still serve and record, but do not train. Tried Songe on another setup? Share the result on the testing board.
🌙 How it works
your agent ──► Songe (127.0.0.1:7613) ──► your model server
│ records what happened
▼
when you are away: a Dream
read lessons ─► practice ─► train ─► check ─► use only if better
| Quick Dream | Deep Dream | |
|---|---|---|
| Takes | a few minutes | an hour or more |
| Does | adds the preferences you stated or showed to the model's instructions | writes practice from your lessons and trains a small adapter on it |
| You see | each new preference, with a before/after where the answer changes | why the new version was or was not used |
| Undo | undo the Quick Dream, or remove one preference | go back to any earlier version |
With auto-dream on, Dreams start when your agent has been quiet for a
while. Or press Dream in the dashboard (songe open). A Dream never just
disappears: it keeps going on the page, or says why it stopped.
How it learns, in more detail
- Reading. Songe reads each conversation for the moments that teach something: your corrections and stated rules, retries, tool errors.
- Practice. For each lesson your model writes short practice scenes in your agent's own tools. Your real conversations are never trained on.
- Training. The model as it was, shown the lesson, is the teacher; the same model with a small adapter, shown only the scene, is the student (on-policy distillation). Earlier skills are rehearsed so they stay.
- Checking. At the moments you reacted to, the version in use and the new one each answer, and a judge compares what they do. The new version is used only if it does clearly better, does not change what already worked, and its answers stay healthy.
Before the first training Songe measures how much your computer can learn at once, and trains on what fits. See DESIGN.md.
⌨️ Commands
| Command | What it does |
|---|---|
songe |
what it is doing |
songe open |
the dashboard in your browser |
songe models |
each model, and whether it dreams on its own |
songe models MODEL on|off |
auto-dream for one model |
songe sleep |
start a Deep Dream now |
songe undo MODEL |
go back one version |
songe pause / songe resume |
stop and restart learning |
songe doctor |
check the setup |
songe forget |
delete recorded conversations (learned versions stay) |
songe uninstall |
stop the service and restore your agent settings |
What it changes on your machine
- Its data folder:
~/Library/Application Support/songeon macOS,$XDG_DATA_HOME/songeon Linux,%LOCALAPPDATA%\songeon Windows. - A
songe/folder in LM Studio's models folder with the current learned version. Your own models are never touched. - One provider entry in opencode and pi, with a backup.
- A background service that starts at login.
Upgrade and remove
uv tool upgrade songe, pipx upgrade songe or brew upgrade songe: the
service moves to the new version by itself in a quiet moment.
songe uninstall stops the service and restores your agent settings; add
--purge to also delete its data and learned versions. Run it before
brew uninstall songe.
🧭 Roadmap and contributing
Where Songe is going: ROADMAP.md. Ideas and bugs are welcome as issues; to contribute, see CONTRIBUTING.md. Security reports: SECURITY.md.
Licensed under Apache-2.0.
Metadata
Release files for songe 0.1.0a14
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| songe-0.1.0a14.tar.gz | 843.8 kB | With provenance | Details for songe-0.1.0a14.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| songe-0.1.0a14-py3-none-any.whl | Python 3 | none | any | Details for songe-0.1.0a14-py3-none-any.whl |
Total release size: 1.4 MB
Release files / songe-0.1.0a14.tar.gz
| Download URL | songe-0.1.0a14.tar.gz |
|---|---|
| Size | 843.8 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
2c15c5d36ec17521655118704267b7ea83989c37e629ba45e326825c93663a2d
|
|
BLAKE2b-256 checksum How to use checksums |
552c402a5a725d71cfa38907761939f880d378a39411185456b8690a91a3e9d4
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 21, 2026.
Transparency logRelease files / songe-0.1.0a14-py3-none-any.whl
| Download URL | songe-0.1.0a14-py3-none-any.whl |
|---|---|
| Size | 507.2 kB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
35632eb2dc54acf1b472634c1cb1ad419515b6832c577db3726f7b5d70484c95
|
|
BLAKE2b-256 checksum How to use checksums |
6a0ea14e88df336dcd0328dd76936f3ef05e995b63a6da1dfc95d201a8a96206
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 21, 2026.
Transparency logSecurity for songe 0.1.0a14
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for songe-0.1.0a14.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 21, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.