Skip to main content
What Monte can see depends on the deployment option: Monte Cloud or BYOC.

​
Data Visibility

Run metadata always goes to the control plane, because run history, monte experiment show, and monte compare read it. Monte never uses a team’s data, traces, or traffic to train or improve any model. With BYOC, traces and job log lines reach Monte only when the team turns them on in the Console under Team → Data Sharing; both are off by default. A job without --remote runs on the machine that runs the CLI, and its data stays there.

​
Follow the Data

With BYOC, data stays in the data plane in the team's account from seed examples to the endpoint; only a closed-weight eval judge calls its provider's API With BYOC, data stays in the data plane in the team's account from seed examples to the endpoint; only a closed-weight eval judge calls its provider's API Seed examples, datasets, and checkpoints stay in artifact storage and on the GPU instances. Each run sends its scores to run history. Requests to a deployment go from the team’s application to the endpoint. Only a closed-weight judge in an eval, or grader code that calls an outside API, sends data out of the data plane. Base weights download once from Hugging Face into artifact storage, and later jobs read that copy; the HF_ENDPOINT secret sets a mirror. Signed job images are copied into the team’s registry, and jobs pull only from it. With BYOC, monte secret set writes the value to the cloud’s secret store with the signed-in cloud credentials, and the control plane stores only the name.

​
Choose a Region

The control plane runs in the EU or the US, chosen when the team is created, and the team’s run metadata stays there. It holds names, configs, statuses, scores, and metrics, and no task text, dataset rows, prompts, or weights. With BYOC, monte cloud connect --region sets one region for each connected cloud. Jobs, storage, and endpoints for that cloud stay in that region. Monte never copies data or weights between connected clouds, and a deployment uses checkpoints from its own cloud.

​
Inference Data

The endpoint keeps no request or response content, and Monte never trains on inference traffic. Request logs stay in artifact storage, which is in the team’s account with BYOC. They hold the time, the deployment, the token counts, and the status code, and never request or response content. They are kept for 30 days. --request-logs off on a deployment keeps none. monte deployment logs prints server logs, which never include prompts or completions.

​
Keep Model Calls Inside the Cloud

The judge model in an eval grader is an open-weight or a closed-weight model. The team chooses. The judge in a training environment, and the generator and the judge in synthetic data, are always open-weight models. Monte calls a provider only for an eval judge named as <provider>/<model>, or from grader code that calls an outside API.
Check the model names and the grader code before a job runs on private data.

​
Limit and Revoke Access

With BYOC, Monte acts only through the operator in the team’s cluster. The operator’s workload identity covers only its namespace, artifact storage, the registry, and the secret store, and the control plane holds no credentials. The Console’s Audit Log records which member started, deployed, or deleted each object, and the cloud provider’s audit log records each action of the operator’s identity. To stop all Monte access at once, remove the operator or its identity. No new jobs start, and running training and eval jobs stop. Running deployments keep serving, but Monte can no longer change them. The data stays in the team’s account.