Integration · ML platform
MLflow + OrchestrAI
Catalog exported 2026-09-02 · MLflow website
Track MLflow experiments and runs from chat, log params and metrics, and move model versions between stages.
OrchestrAI exposes 12 MLflow operations: 11 are low-risk (read-only or low-impact), and 1 create or modify resources and run only after you confirm the plan.
What teams use it for
Teams point OrchestrAI at a tracking server to answer questions like which run in the churn-v3 experiment scored highest, then register the winner and promote it to Staging without opening the UI. The agent can also open a run, write parameters and metrics into it, and close it, which is handy when a training job runs elsewhere and you want the result recorded. Stage transitions are the only medium-risk step; everything else is read or append. There is no operation to delete experiments, runs, or model versions, and model artifacts themselves are not uploaded from chat, so use the MLflow client for those.
Every MLflow operation, with its risk level
| Operation | What it does | Risk | Step-level approval |
|---|---|---|---|
Create MLflow Experiment |
Create an experiment | Low risk | No |
End MLflow Run |
End current run | Low risk | No |
Get MLflow Experiment |
Get experiment details | Low risk | No |
Get MLflow Run |
Get run details | Low risk | No |
List MLflow Experiments |
List experiments | Low risk | No |
List MLflow Registered Models |
List registered models | Low risk | No |
List MLflow Runs |
List runs in experiment | Low risk | No |
Log MLflow Metrics |
Log metrics to run | Low risk | No |
Log MLflow Parameters |
Log parameters to run | Low risk | No |
Register MLflow Model |
Register a model | Low risk | No |
Start MLflow Run |
Start a new run | Low risk | No |
Transition MLflow Model Stage |
Transition model version stage | Creates resources | No |
Risk tiers come from the catalog: low is read-only or low-impact, medium creates resources and is reversible, high modifies existing resources, destructive may lose data. Every plan that creates or changes resources is shown with its cost estimate and waits for your confirmation. Operations marked with a step-level approval pause again on their own step. Destructive operations require a typed risk phrase.
What you connect
A MLflow credential (stored as mlflow).
Connected-service tokens are envelope-encrypted with a per-record key wrapped by a cloud KMS.
Prompts that work
- List the runs in the fraud-detector experiment and show me the one with the lowest validation loss
- Register the model from run 8f3a as fraud-detector and transition version 4 to Production
- Start a run in experiment pricing-lgbm, log learning_rate=0.05 and num_leaves=63, then log auc=0.91 and end the run
Before anything runs
Every mutation shows its plan, cost estimate, and blast radius, then waits for your confirmation. Destructive operations require a typed risk phrase. Credentials are minted per run through OIDC federation and discarded afterward; nothing you create here is invisible later, because every resource lands in the desired-state ledger where drift is detected and can be converged. Details on the security page.
Frequently asked questions
- Can OrchestrAI promote an MLflow model version to Production?
- Yes, mlflow_transition_model_stage changes a registered model version's stage and is rated medium risk because it changes what downstream consumers load. Registering a model in the first place is a low-risk operation.
- Does OrchestrAI need write access to my MLflow tracking server?
- Only if you want it to create experiments, start runs, or log parameters and metrics. Listing experiments, runs, and registered models works with read access alone, and none of the MLflow operations remove data.
- How does OrchestrAI authenticate to MLflow?
- You add a MLflow credential once in the connections screen. It is envelope-encrypted with a per-record key wrapped by a cloud KMS and is only decrypted inside the run that needs it.
Related integrations
Try it on your own account
Connect your cloud read-only and see your resources, drift, and costs before anything runs. $5 minimum to start. Unused credits refunded in your first 14 days.
Unused credits refunded in your first 14 days.