A little clarity.
A lot less setup.

A Model-as-a-Service platform. You upload a trained model file, or deploy it from your notebook with the Python client, and get back a live, key-protected HTTPS endpoint you can call from anywhere. You never deal with servers, Docker or pip.

Pickle and joblib files from scikit-learn, XGBoost, LightGBM and CatBoost, XGBoost's own .json and .ubj files, ONNX files and TensorFlow Lite files, up to 50 MB. With the Python client you can also deploy Keras and PyTorch models, converted for you, and small Hugging Face text models.

MLGatee never loads your file on its own servers. Loading a pickle can run code hidden inside it, so the inspector only reads the raw bytes to find the libraries and versions it needs, and refuses files that would run commands when loaded. Your model then runs in its own isolated service, shared with no other model.

The message says why in plain words. Common reasons: the model uses a class defined in your notebook or your own .py file, the file would run commands when loaded, the framework isn't supported yet, or a joblib file uses bz2, xz or lz4 compression (zlib and gzip work).

Typically 1.5 to 4 minutes from upload to a live endpoint. CatBoost models take a little longer because the library is larger.

On the free plan a model sleeps after 15 minutes without calls. The next call wakes it, which takes about a minute. Paid plans keep models awake.

Upload a new file for the same model, or deploy the same name again from Python. It becomes the next version at the same address with the same key. The current version keeps answering while the new one builds, and if the build fails the current version stays live.

Yes. The files of your three newest versions are kept, so you can roll back to any of them in one click or one line of Python. A rollback becomes a new version built from the earlier file.

Counts and timings only: calls, errors, answer times and when the model was last called, per hour and per version. Inputs, outputs, keys and headers are never recorded.

Create a new key on the model's page. The model rebuilds with it, and the old key keeps working until the rebuild is live, so nothing breaks in between. Keys are shown once and stored only as a hash.

The Python client converts Keras models to TensorFlow Lite and PyTorch models to ONNX in your own Python, and checks the result gives the same answers before uploading. Small Hugging Face text models are packed into one ONNX file with their tokenizer. Native TensorFlow, PyTorch and full Hugging Face models need more memory and are planned for the Large machine add-on.

Yes. Sign in at app.mlgatee.com with GitHub, Google or an email link. The free plan runs one model, and there’s no card to enter. The Python client is open source under the Apache-2.0 licence.