mlgatee

Deploy a trained model to MLGatee from a notebook, a script or the terminal, and get a live, key-protected HTTPS endpoint.

%pip install -q mlgatee

import mlgatee
mlgatee.login()                                  # paste an access token (hidden)
dep = mlgatee.deploy(model, name="churn-model")  # about 1.5 to 3 minutes
dep.predict([[5.1, 3.5, 1.4, 0.2]])

✓ saved with joblib (3 KB)
✓ uploaded
✓ checked without running it: numpy==2.5.3, scikit-learn==1.9.1

  Endpoint key (shown once, also in .key): mlg_live_…
  Save it somewhere safe: only its last 4 characters are kept.

… Installing libraries and downloading the model
… Starting the model server and loading the model
✓ live at https://mlg-churn-model-….onrender.com

Get an access token on app.mlgatee.com → Settings → Access tokens. MLGatee is invite-only during the beta.

How it works

The model is saved with joblib in your own Python, then goes through exactly the same steps as an upload on the web app: MLGatee reads the file's bytes (it never runs your code), picks the exact libraries, and gives the model its own server on Render.

Deploying the same name again uploads a new version: same address, same key. The current version keeps answering until the new one is live; if the new build fails, it keeps answering.

Rolling back builds a new version from an earlier version's file: rolling v3 back to v2 creates v4 with v2's file. The 3 newest versions can be rolled back to.

Python

Errors are mlgatee.MLGateeError with a plain message and a stable .code.

Terminal

pip install mlgatee
mlgatee login
mlgatee deploy model.pkl --name churn-model
mlgatee models
mlgatee versions churn-model
mlgatee rollback churn-model 2
mlgatee plan
mlgatee resume churn-model
mlgatee delete churn-model --yes

Monitoring

mlgatee.usage("churn-model")              # last 24 hours; "7d" for a week
print(mlgatee.usage_text(mlgatee.usage("churn-model")))
dep.usage()                               # the same, from a Deployment

mlgatee usage churn-model --window 7d

Calls, errors (bad requests and server errors), answer times (p50, p95, max), rows predicted and when the model was last called, overall and per version: handy before deciding on a rollback. MLGatee records counts and timings only, never inputs or outputs. A model deployed before monitoring existed starts reporting after its next build.

Keras and PyTorch

Pass the model itself: it's converted here, in your Python, to a format MLGatee's free machines run (TensorFlow Lite for Keras, ONNX for PyTorch), checked against the original, then uploaded.

mlgatee.deploy(keras_model, name="my-keras-model")                      # -> TensorFlow Lite
mlgatee.deploy(keras_model, name="my-keras-model", quantize=True)       # 8-bit weights, ~4x smaller
mlgatee.deploy(torch_model, name="my-torch-model", example_input=X[:5], task="classifier")  # -> ONNX

• Keras needs TensorFlow (the backend you trained with), PyTorch needs torch and onnx (pip install "mlgatee[torch]" installs the extras). With example_input, the converted model is compared with the original before uploading (needs ai-edge-litert / onnxruntime).
• task="classifier": /predict returns the class and /predict_proba the probabilities (softmax of the scores). Without it, the model's outputs are returned as they are.
• Custom or Lambda layers, and operations TFLite/ONNX don't have, are refused with the converter's own message. Full TensorFlow and PyTorch will come on the paid plan.
• Already have a file? mlgatee.deploy("model.tflite", ...) or "model.onnx" works too.

Hugging Face text models

Small text models (MiniLM-sized: up to about 25 MB with 8-bit weights) run on the free plan:

mlgatee.deploy_hf("sentence-transformers/all-MiniLM-L6-v2", name="embeddings")
mlgatee.deploy_hf("philschmid/tiny-bert-sst2-distilled", name="sentiment")

Callers send plain texts: {"input_data": ["great product!"]}. A classifier answers label names (and /predict_proba the probabilities); an embeddings model answers one normalised vector per text. The model is exported to ONNX here, with 8-bit weights, checked against the original, and packed with its tokenizer into one file (pip install "mlgatee[huggingface]"). Bigger models (DistilBERT and up) don't fit the 50 MB limit; text generation comes later.

Good to know

• Supported during the beta: scikit-learn, XGBoost, LightGBM and CatBoost, up to 50 MB, one model per person, 5 builds a day.
• A model that uses a class defined in your notebook can't be loaded on a server. The client catches this before uploading and says which class to move into a .py file.
• The endpoint key is shown once. Keep it out of notebooks you share.
• The free instance sleeps after 15 idle minutes; the next request takes about a minute.

Development

pip install -e ".[dev]"
python -m pytest -q

Self-hosting MLGatee? Point the client at your project with MLGATEE_URL and MLGATEE_PUBLISHABLE_KEY.

Apache-2.0.