From one GPU to a Ray cluster, without leaving your editor.

Python

# Inside a workspace: plain Ray, nothing Fantasti-specific
import ray
ray.init()  # connects to this workspace's Ray cluster

ds = ray.data.read_parquet("s3://acme-data/prompts/")
ds = ds.map_batches(  # workers scale up from 0
    Embedder, concurrency=8, num_gpus=1, batch_size=256,
)
ds.write_parquet("s3://acme-data/embeddings/")

head      ready    16 vCPU · on-demand
workers   0 → 1    B200:8 · spot · map_batches asked for 8 GPUs
workers   1 → 0    released when the work is done

CLI

fantasti workspace create rl-dev --gpu H100:1 \
  --image ghcr.io/acme/train:latest --idle-stop 45m
fantasti workspace open rl-dev --ide cursor          # or: vscode | jupyter | ssh

# Scale out: the CPU head keeps your editor and files; B200 workers scale from zero
fantasti workspace scale rl-dev --workers B200:8 --min 0 --max 4 \
  --spot --max-price follow

# Ship the same code as a job (inherits image, env, secrets, files)
fantasti job submit --from-workspace rl-dev -- uv run train.py --epochs 3

MCP

{
  "mcpServers": {
    "fantasti": { "type": "http", "url": "https://mcp.fantasti.ai/mcp" }
  }
}

workspaces_list
workspaces_start
workspaces_stop    asks for confirmation

YAML

# workspace.yaml  (fx.workspaces.create(spec="workspace.yaml"))
name: rl-dev
image: ghcr.io/acme/train:latest
head: { cpu: 16, capacity: on_demand }   # editor, Ray head, files
workers:
  - gpu: B200:8
    min: 0
    max: 4
    capacity: { mode: spot, max_price: follow }   # or a USD limit per GPU-hour
env: { WANDB_MODE: offline }
secrets: [HF_TOKEN]
volumes: ["s3://acme-data:/mnt/data:ro"]
idle_stop: 45m
mlflow: default                         # sets MLFLOW_TRACKING_URI on every node