| name | vessl-cloud-onboarding |
| description | Use when signing up or 가입 for current VESSL Cloud, installing or authenticating vesslctl, checking organizations and teams, or preparing a first billable GPU Workspace, Job, or 기본 세팅. |
VESSL Cloud Onboarding
Overview
Use only the current VESSL Cloud at https://cloud.vessl.ai and the vesslctl CLI. Separate free/read-only setup verification from billable compute, storage, and Job creation. Read the official workflow and command reference when acting. For autoresearch, also read the official VESSL repositories and defer to the pinned vessl-cloud-cookbook/autoresearch recipe.
Preflight
Classify the target as signup, CLI verification, cost review, or paid tutorial. Confirm the current product URL and whether vesslctl is installed without reading credentials or ~/.config/vesslctl/config.yaml.
Product and Secret Boundary
Do not mix the current Cloud product with legacy VESSL MLOps, its vessl Python CLI, or its Organization/Project model. Current Cloud uses Organization → Team → Workspace/Job.
Never ask for or inspect passwords, MFA codes, OAuth tokens, access-token values, cookies, browser storage, or ~/.config/vesslctl/config.yaml. The user completes credentials, OAuth/SAML, CAPTCHA, email verification, legal acceptance, age confirmation, and final signup submission.
Workflow
Phase 1 — Signup and Free Setup Verification
- Open the official signup page with the available Browser skill and inspect visible state. If browser control is unavailable, give the official link and manual checkpoints.
- Hand sensitive signup and Organization creation submission to the user. Explain that the creator may receive administrative responsibility.
- Before CLI installation, inspect the official install instructions, explain the binary/PATH change, and obtain explicit confirmation.
- Enforce browser authentication with
vesslctl auth login --web; the user completes the sensitive browser step. Do not fall back to terminal password entry.
- Verify only with non-secret, read-only commands:
vesslctl auth status
vesslctl config show
vesslctl billing show
vesslctl org list
vesslctl team list
vesslctl cluster list
vesslctl resource-spec list --usable-only
Completing these checks is a valid stopping point. Do not imply that a paid tutorial is required for setup success.
Phase 2 — Billable Compute and Storage Gate
Copy assets/workspace-cost-card.md and fill it from current official UI/CLI values. Before vesslctl volume create, vesslctl workspace create, or vesslctl job create, show Organization/Team, cluster/region, resource spec and GPU count, availability, hourly price, current credit, container image, storage capacity/rate/estimated duration and cost, mounts, exposed ports, expected duration, timeout behavior, and cleanup plan. Obtain explicit confirmation for that exact resource.
After approval, create the smallest suitable resource, run python -c 'print("Hello, VESSL Cloud!")', verify logs/output, and return immediately to cleanup.
Official autoresearch profile
For auto-research, use vessl-ai/vessl-cloud-cookbook/autoresearch pinned at 97a0af14b0acae042162b1f70f17fbe2d570afa2 as the execution source of truth. Its batch-job/prep.sh and batch-job/submit.sh create billable Jobs and mount an object volume. Use a participant-owned writable fork and an exact live single-A100 resource spec from vesslctl resource-spec list --usable-only -o json. Export that slug explicitly; never accept the recipe's H100 default or substitute another GPU.
Keep the official benchmark and batch-job scripts unchanged. The Ralphthon overlay bounds the recipe to one baseline, at most three sequential candidates, and one confirmation; it does not use async fan-out or an unbounded loop. The W&B recorder runs locally after vesslctl job logs is fetched, so no W&B API key is placed in VESSL.
Cleanup Gate
Pause stops Workspace compute but can preserve state and leave storage charges. Terminate is destructive and stops the Workspace or Job; object volumes and other separately managed storage may remain. Show the consequence and obtain explicit confirmation immediately before either action unless the exact timeout termination was already part of the approved creation card. Never rely on credit exhaustion as automatic cleanup.
AUTORESEARCH_TIMEOUT_S stops only the cookbook's local polling. The remote Job keeps running and billing unless it has independently reached a terminal state. On timeout, inspect vesslctl job show <slug> -o json, show the live state, execute only the approved vesslctl job terminate <slug> action, and recheck terminal state and storage before claiming cleanup.
Completion Levels
| Level | Evidence |
|---|
| Account ready | User confirms verified signup and Organization |
| CLI ready | Auth/config/org/team checks succeed without revealing tokens |
| Cost reviewed | Live cost card is complete; no compute created |
| Tutorial complete | Approved resource prints Hello, VESSL Cloud and reaches a verified cleanup state |
Verification
- Verify auth/config/org/team/cluster/resource-spec state without exposing secrets.
- For compute, verify the exact approved resource, Job/Workspace output and terminal state, live hourly exposure, current billing state, and final Pause/Terminate/storage state.
- Never claim all charges ended until compute and any separate persistent storage are both checked.
Output
Return the completion level, verified evidence, waiting-for-user action, exact planned or completed side effect, hourly exposure, current billing/credit state, and cleanup state.
Next Steps
- CLI ready: stop without cost, or prepare the live cost card.
- Cost approved: create one minimal resource and run Hello VESSL.
- Tutorial complete: obtain and execute a separate cleanup decision.
Common Mistakes
- Using
vessl configure → stop; that is the legacy product.
- “Cheapest GPU” without live inventory → compare
resource-spec list --usable-only and prices.
- Creating before cleanup planning → fill the complete cost card first.
- Claiming Pause is free → report continuing storage or related charges.
- Treating a polling timeout as cleanup → inspect and terminate the remote Job only with the approved action.
- Using the cookbook's default H100 for an A100 request → export the exact live A100 spec before creation.
