Reference

CLI reference

gft is the command-line tool for GFT Studio. Run gft --help, or gft <command> --help, to see every option. Install it with the Quickstart.

Settings#

Name	Meaning
`--json`	Print machine-readable JSON. Put it before the command: `gft --json run list`.
`--yes`, `-y`	Skip the cost confirmation on billable commands.
`GFT_API_URL`	API origin. Use `https://studio.g-ftech.com`; the default is `http://127.0.0.1:8787`.
`GFT_API_KEY`	API key for scripts. Takes priority over the saved key.
`GFT_STUDIO_API_TOKEN`	Token for a local Studio operator API, used by `--local` and `gft deploy bundle`.
`~/.config/gft/config.toml`	Where `gft login` saves your key, readable only by you.

Account#

Command	What it does
`gft login`	Check an API key and save it.
`gft logout`	Delete the saved key. The key stays valid on the server.
`gft whoami`	Show the current user and organization.
`gft doctor`	Check API health and authentication.

Gyms#

Command	What it does
`gft env list`	Search and filter gyms: `--ready`, `--evaluation-ready`, `--search`, `--family`, `--provider`, `--source`, `--owner`, `--status`, `--limit`.
`gft env info <slug>`	Show one gym, its readiness and schemas.
`gft env add <manifest>`	Publish a gym whose implementation is installed on the Hub.
`gft env push <wheel> --manifest <file>`	Upload a wheel and build its sandbox. Add `--follow` to wait.
`gft env connections`	List private gym connections.
`gft env connect --manifest <file>`	Publish a gym backed by `--image` or `--connection`.

Training and evaluation#

Command	What it does
`gft train start <slug> --model <model>`	Start a GRPO training run. See Train with RL.
`gft eval start <slug> --model <model>`	Start a held-out evaluation. See Evaluate a model.
`gft run list`	List runs. Filter with `--kind training` or `--kind evaluation`.
`gft run show <id>`	Show one run or evaluation.
`gft run watch <id>`	Follow a run until it finishes. Add `--exit-status` for scripts.
`gft run cancel <id>`	Cancel a run or a queued evaluation.

Models and inference#

Command	What it does
`gft deploy artifacts`	List promoted adapters you can deploy.
`gft deploy start <artifact-id>`	Provision, canary-test and activate an adapter.
`gft deploy list`, `show`, `token`, `remove`	Manage adapter deployments.
`gft deploy machines <model>`	List cloud machine tariffs for a model.
`gft deploy cloud <model>`	Deploy a model on a cloud machine.
`gft deploy bundle`	Prepare a Helm bundle for your own cluster.
`gft chat <model> <prompt>`	Send a prompt. Add `--stream` to print tokens as they arrive.
`gft bench <model>`	Measure latency and throughput.

Containers#

Command	What it does
`gft image add <reference>`	Register an image by digest. Add `--private` for a private registry.
`gft image list`	List registered images.
`gft container capabilities`	Show tariffs, modes and whether containers are available.
`gft container run <image-id>`	Start a job or a service.
`gft container list`, `get`, `logs`, `outputs`, `stop`	Inspect, download results from and stop runs.
`gft container request <run-id>`	Send one HTTP request to a running service.
`gft container shell <run-id>`	Open an interactive terminal.

Python SDK#

The same package provides a Python client. Client() reads GFT_API_URL and GFT_API_KEY:

python

from gft_client import Client

with Client() as client:
    print(client.me())
    for environment in client.environments.list():
        print(environment)

The client exposes environments, runs, benchmarks, deployments, chat, images and containers. It retries reads and idempotent writes automatically after network errors and 429, 502, 503 or 504 responses.