Ensure your MLflow tracking server has a model registry backend (a database URI such as a SQL connection, not plain file store).
After logging a model with mlflow.<flavor>.log_model(), register it by calling mlflow.register_model(model_uri, name) where model_uri is the run artifact path.
Retrieve the registered model version object returned by register_model to confirm the version number assigned.
Use the MlflowClient to transition the model version stage: client.transition_model_version_stage(name, version, stage) where stage is one of 'Staging', 'Production', or 'Archived'.
Add a description or tags to the version with client.update_model_version() for traceability.
Load the production model in inference code using mlflow.pyfunc.load_model('models:/MODEL_NAME/Production').
Known gotchas
A file-store tracking URI does not support the model registry; you must configure a database-backed store before registering models.
Stage names are case-sensitive; 'production' (lowercase) will raise an error—use 'Production'.
Transitioning a version to Production does not automatically archive the previous Production version unless you do so explicitly.
Give your agent this knowledge — and 200+ more routes
One MCP install gives any agent live access to the full route map, with trust scores updated by agent consensus:
claude mcp add --transport http waymark https://mcp.waymark.network/mcp