Setup - Monte Platform

These commands prepare an account before the first run. Conventions covers addresses, shared flags, and exit codes.

Command	Description
`monte login`	Sign in
`monte logout`	Remove the sign-in token from this machine
`monte config`	Print the resolved configuration
`monte --version`	Print the installed version
`monte cloud connect`	Connect a cloud account for BYOC
`monte cloud list`	List connected clouds
`monte cloud show`	Show one connected cloud
`monte cloud disconnect`	Remove a connected cloud
`monte secret set`	Store a secret that jobs read
`monte secret list`	List secret names
`monte secret delete`	Remove a secret
`monte instance up`	Start a GPU instance
`monte instance status`	List GPU instances
`monte instance down`	Stop a GPU instance
`monte mcp`	Print the MCP server setup

monte login [--no-browser]

Open a browser to sign in, and store a token on this machine. The same account signs in to the Console and the MCP server.

Flag	Type	Default	Description
`--no-browser`	flag	off	Print a URL and a one-time code to open on any device

monte login

Signed in as [email protected] (team helpdesk).

monte logout

monte logout

Remove the sign-in token from this machine. Running jobs and inference API keys do not change.

monte logout

Signed out [email protected]. Removed ~/.config/monte/token.

monte config

monte config [--json]

Print the resolved configuration: the account, the data root, the artifact bucket, the connected clouds, and the training configs.

Flag	Type	Default	Description
`--json`	flag	off	Print JSON on stdout

monte config

account        [email protected] (team helpdesk)
data root      ~/monte-data
artifacts      gs://helpdesk-monte-artifacts
clouds         gcp (us-central1)
configs        triage-sft, triage-dpo, triage-grpo, triage-distill

MONTE_CONFIGS adds a directory of training configs. MONTE_GYM_ROOTS sets where monte env create writes and where monte env add reads.

monte —version

monte --version

Print the installed version.

monte --version

monte 0.14.2

monte cloud connect

monte cloud connect <azure|aws|gcp> --region <region> [--name <name>]
  [--cluster <name> --namespace <namespace>]
  [--subscription <id>] [--resource-group <name>]   # Azure
  [--account <id>]                                  # AWS
  [--project <id>]                                  # GCP

Connect a cloud account for BYOC: jobs, artifact storage, and deployments then run in it. The command creates the resources that Connect a Cloud for BYOC lists, and installs the Monte operator with one identity scoped to its namespace and the artifact bucket. It uses the signed-in gcloud, az, or aws session, and that account needs the cloud roles listed there.

Flag	Type	Default	Description
`<azure\|aws\|gcp>`	argument	required	The cloud provider
`--region`	string	required	The one region for jobs, storage, and endpoints in this cloud
`--name`	string	the provider name	The name that `--cloud` flags and `monte cloud show` use
`--cluster`	string	none	An existing Kubernetes cluster. The command installs the operator in it and creates no cluster or node pool. It still creates artifact storage, the registry, and the secret store.
`--namespace`	string	required with `--cluster`	The namespace for the operator and its jobs
`--subscription`	string	required on Azure	The Azure subscription ID
`--resource-group`	string	`monte`	The Azure resource group that holds the resources
`--account`	string	required on AWS	The AWS account ID
`--project`	string	required on GCP	The GCP project ID
`--yes`	flag	off	Skip the prompt that lists the resources to create
`--json`	flag	off	Print JSON on stdout

monte cloud connect gcp --project <project-id> --region us-central1

This creates in project <project-id> (us-central1):
  GKE cluster monte-gke, GPU node pool gpu-h100 (0 nodes)
  bucket helpdesk-monte-artifacts
  operator in namespace monte, identity monte-operator
  ...
create? [y/N] y
Connected cloud 'gcp' (us-central1). The operator connected outbound. Check the GPU quota of the node pool before the first job.

In an existing cluster:

monte cloud connect azure --subscription <subscription-id> --resource-group monte \
  --region eastus --cluster <name> --namespace monte

Error	Cause
`E_STORAGE`	The account cannot create the bucket
`E_REFUSAL`	A cloud with that name is already connected

monte cloud list

monte cloud list [--json]

List the connected clouds of the team.

Flag	Type	Default	Description
`--json`	flag	off	Print JSON on stdout

monte cloud list

NAME    PROVIDER  REGION         DEPLOYMENTS
gcp     gcp       us-central1   3

monte cloud show

monte cloud show <name> [--json]

Show one connected cloud: each resource, its region, and the scope of the operator’s identity.

Flag	Type	Default	Description
`<name>`	argument	required	The cloud name
`--json`	flag	off	Print JSON on stdout

monte cloud show gcp

cloud      gcp (us-central1)
operator   monte-operator in namespace monte, connected outbound
identity   monte-operator  scope: namespace monte on monte-gke, bucket helpdesk-monte-artifacts, registry, secret store
compute    monte-gke  node pool gpu-h100  0/8 nodes
storage    helpdesk-monte-artifacts

monte cloud disconnect

monte cloud disconnect <name> [--yes] [--json]

Remove the operator and its identity. The resources and the data stay in the team’s account.

Flag	Type	Default	Description
`<name>`	argument	required	The cloud name
`--yes`	flag	off	Skip the prompt
`--json`	flag	off	Print JSON on stdout

monte cloud disconnect gcp --yes

Disconnected cloud 'gcp'. The resources stay in project <project-id>.

Error	Cause
`E_IN_USE`	A deployment runs in the cloud. Delete it with `monte deployment delete` first.

monte secret set

monte secret set <NAME> [--cloud <name>] [--json]

Store a secret for the team. Every job of the team reads it as an environment variable with the same name. A grader reads the API key of a judge model this way. The value comes from a hidden prompt, or from stdin. No flag takes the value, so it never goes into shell history. A second set of the same name replaces the value.

Flag	Type	Default	Description
`<NAME>`	argument	required	Uppercase letters, digits, and `_`
`--cloud`	string	the only connected cloud	Store the secret in the secret store of that cloud. On Monte Cloud, Monte stores it.
`--json`	flag	off	Print JSON on stdout

monte secret set JUDGE_API_KEY

value for JUDGE_API_KEY: ********
Stored JUDGE_API_KEY in gcp (Secret Manager).

monte secret list

monte secret list [--json]

List the secret names of the team, with the store and the time of the last change. Values never print.

Flag	Type	Default	Description
`--json`	flag	off	Print JSON on stdout

monte secret list

NAME            STORE   UPDATED                BY
HF_TOKEN        gcp     2026-10-02 09:14 UTC   [email protected]
JUDGE_API_KEY   gcp     2026-10-27 16:40 UTC   [email protected]

monte secret delete

monte secret delete <NAME> [--yes] [--json]

Remove a secret. Jobs that start after the delete do not get the variable. Running jobs keep the value.

Flag	Type	Default	Description
`<NAME>`	argument	required	The secret name
`--yes`	flag	off	Skip the prompt
`--json`	flag	off	Print JSON on stdout

monte secret delete HF_TOKEN --yes

Deleted HF_TOKEN from gcp.

monte instance up

monte instance up [--type <gpu>x<n>] [--cold] [--keep-alive] [--json]

Start a GPU instance that stays up across several jobs. --remote starts an instance on its own when none is free. The 24-hour maximum lifetime applies only to instances that monte instance up starts. A --remote job runs until it finishes.

Flag	Type	Default	Description
`--type`	string	`h100x1`	The GPU type and the GPU count: `l4`, `l40s`, `a100-80gb`, `h100`, or `h200`, then `x1`, `x2`, `x4`, or `x8`
`--cold`	flag	off	Start with an empty cache, and fetch environments, datasets, and weights from artifact storage
`--keep-alive`	flag	off	Turn off the idle shutdown. The 24-hour maximum lifetime still applies.
`--json`	flag	off	Print JSON on stdout

monte instance up --type h100x1

Started gpu-4f2a (h100x1, us-central1). Ready in 3m12s.
Idle shutdown after 30m. Maximum lifetime 24h.

Error	Cause
`E_INSTANCE`	No instance of that type is available, or the GPU quota is full

monte instance status

monte instance status [--json]

List the GPU instances of the account, with the age, the job, and the time to shutdown.

Flag	Type	Default	Description
`--json`	flag	off	Print JSON on stdout

monte instance status

ID         TYPE          AGE     JOB                       SHUTDOWN
gpu-4f2a   h100x1        1h05m   support-triage/train-07   30m after the job
gpu-91c0   h100x1        12m     none                      in 18m

monte instance down

monte instance down <id|name> [--yes] [--json]

Stop one GPU instance. A job on it stops, and its run reads as abandoned.

Flag	Type	Default	Description
`<id\|name>`	argument	required	The instance ID or name
`--yes`	flag	off	Skip the prompt when a job runs on the instance
`--json`	flag	off	Print JSON on stdout

monte instance down gpu-91c0

Stopped gpu-91c0 after 12m.

monte mcp

monte mcp [--agent claude|cursor|codex]

Print the MCP server URL and a setup snippet for each supported coding agent.

Flag	Type	Default	Description
`--agent`	string	all three	Print only the snippet for one agent

monte mcp --agent claude

MCP server  https://mcp.trymonte.ai/mcp

Claude Code:
  claude mcp add --transport http monte https://mcp.trymonte.ai/mcp

Conventions

Address a Run or a Checkpoint

Run IDs count up separately for each kind: train-NN and eval-NN.
A run is <experiment>/<run>, for example support-triage/train-07. The run half is the run ID or an alias: eval@N or train@N, the succeeded run at absolute step N.
In a command that already names the experiment, a checkpoint is <run>/<step>: monte eval --checkpoint train-07/200 and monte train --parent train-07/200. <step> is absolute, counted from the base model. Aliases do not resolve here.
In a command that names no experiment, a checkpoint is <experiment>/<run>/<step>, for example support-triage/train-07/200. The short form <experiment>/<run> names the checkpoint that the run saved at its last step.
An environment or a dataset is <name> or <name>@<version>, for example support-escalation@2. A bare name means the newest version.
Two baselines share the alias eval@0. monte compare takes the one on the same eval as the other run. Other commands refuse the alias and list the IDs.

Shared Flags

Flag	Commands	Effect
`--json`	Every command except `login`, `logout`, `mcp`, and `run logs`	Print one JSON document on stdout. Notes go to stderr. `monte run logs` prints JSON lines already.
`--yes`	Every command that asks before it acts	Skip the prompt. Without a terminal and without `--yes`, the command refuses with `E_CONFIRM`.
`--dry-run`	`eval`, `train`, `train resume`, `grade`	Resolve the settings, print them, and start nothing
`--remote`	`eval`, `train`, `train resume`, `grade`, `dataset generate`	Submit the job to a GPU instance and return. On `experiment list`, read artifact storage.
`--push`	`eval`, `train`, `train resume`, `grade`	Upload results after each state change. On a training run, also push the checkpoint at the end.
`--smoke`	`eval`, `train`	A 3-step training run or a 3-task eval that never counts as a result

--confirm <experiment> replaces --yes on experiment delete. Its value is the experiment name, typed out.

Exit Codes

Every refusal or failure prints CODE: cause (fix) on stderr. Codes 3 and higher are refusals or failures.

Code	Name	When
0	success
1	aborted	Ctrl+C at the prompt, or a broken pipe
2	usage	A bad flag, a missing argument, or two flags that exclude each other
3	`E_REFUSAL`, `E_NO_BASELINE`	A refusal with no more specific code, or a training run before each eval has a baseline
5	`E_NOT_COMPARABLE`	A comparison or a regrade across runs whose tasks, eval settings, or base model differ
6	`E_PREFLIGHT`	A preflight check failed: the run lock, a changed eval environment, disk space, or the GPU
7	`E_CONFIRM`	The prompt was declined or could not show, or a `--confirm` value does not match
8	`E_EXPERIMENT_EXISTS`	`experiment create` on a name that exists
9	`E_NOT_FOUND`	No such object
10	`E_NOT_RUNNING`	`run stop` on a run that is not running
11	`E_CORRUPT_STATE`	Stored state fails validation, or a stored version fails its hash
12	`E_SERVING`	The model server produced no score
13	`E_VERIFICATION`	A finished run or a checkpoint file failed its check
14	`E_STORAGE`	A bucket is unreachable
15	`E_INSTANCE`	The CLI could not get or operate a GPU instance
16	`E_REMOTE`	A `--remote` job could not be submitted
18	`E_NOT_SIGNED_IN`	No sign-in token, or a token that is not valid
19	`E_NOT_A_MEMBER`	Signed in, but not a member of the team that owns the records
20	`E_IN_USE`	A delete or disconnect of an object that another record still uses

​monte login

​monte logout

​monte config

​monte —version

​monte cloud connect

​monte cloud list

​monte cloud show

​monte cloud disconnect

​monte secret set

​monte secret list

​monte secret delete

​monte instance up

​monte instance status

​monte instance down

​monte mcp

​Conventions

​Address a Run or a Checkpoint

​Shared Flags

​Exit Codes

monte login