Get started

Deploy models and manage resources from your terminal.

You can use the Together CLI to manage your Together AI resources from your terminal, or using automated systems. Use the Together CLI to deploy endpoints, fine-tune models, manage your GPU clusters, and more.

Install the Together CLI

Requires Python 3.10+.

Run the following command to install the CLI:

# Install
uv tool install together[cli]

# List commands
tg --help

The [cli] extra target includes additional dependencies used by the CLI exclusively. The extra target prevents the Python SDK from being bloated with dependencies it does not use.

Update the CLI

# Upgrade
uv tool upgrade together[cli]

# run
tg --help

Authenticate

The CLI relies on the TOGETHER_API_KEY environment variable being set to your account’s API token to authenticate requests. You can find your API token in your account settings.

To create an environment variable in the current shell, run:

Shell

export TOGETHER_API_KEY=xxxxx

You can also add it to your shell’s global configuration so all new sessions can access it. Different shells have different semantics for setting global environment variables, so see your preferred shell’s documentation to learn more.

Use the CLI in a CI/CD Environment

uvx is a helper utility from uv that downloads and runs a Python binary without a separate install step. In a CI/CD environment, invoke the CLI directly with uvx:

uvx together[cli] COMMAND [ARGS]...

To authenticate requests the environment variable is recommended for CI/CD because it avoids exposing the token in command-line arguments, which can be visible in process lists and logs. If both are provided, the —api-key flag takes precedence over the environment variable.

Available Commands

models

View Together models and upload your own.

tg models list
tg models upload

Learn more about the models command.

endpoints

Manage your models on your own custom endpoints for improved reliability at scale.

tg endpoints create
tg endpoints start endpoint-id
tg endpoints stop endpoint-id
tg endpoints hardware --available --model my-uploaded-model

Learn more about the endpoints command.

files

Upload and manage datasets for use in fine-tuning, evals, and batch inference.

tg files check ./dataset.jsonl
tg files upload ./eval-data.jsonl --purpose eval
tg files list

Learn more about the files command.

fine-tuning

Fine tune custom models.

tg fine-tuning create
tg fine-tuning list-checkpoints
tg fine-tuning download

# Shorthand alias
tg ft create

Learn more about the fine-tuning command.

evals

Manage model evaluation jobs.

tg evals create 
tg evals status [EVAL_ID]

Learn more about the evals command.

clusters (beta)

Reserve, manage, and interact with gpu clusters.

tg beta clusters create
tg beta clusters get-credentials [CLUSTER_ID]
tg beta clusters list-regions

Learn more about the clusters command.

jig (beta)

Build, deploy and manage dedicated containers.

tg beta jig init
tg beta jig deploy
tg beta jig secrets set HF_TOKEN $HF_TOKEN

Learn more about the jig command.

Beta commands

The beta namespace provides access to experimental features and new capabilities before they become part of the standard API.

Features in the beta namespace are largely considered stable. However, these features are subject to change and may be modified or removed in future releases.

Command aliases

Several commands and subcommands have shorthand aliases for faster typing:

Alias	Full command	Available on
ft	fine-tuning	tg ft … (e.g. tg ft create, tg ft ls)
ls	list	tg files ls, tg fine-tuning ls, tg models ls, tg endpoints ls, tg evals ls, tg beta clusters ls, tg beta clusters storage ls, tg beta jig ls, tg beta jig secrets ls, tg beta jig volumes ls
-c	create	tg fine-tuning -c, tg endpoints -c, tg evals -c, tg beta clusters -c, tg beta clusters storage -c
-d	delete	tg files -d, tg fine-tuning -d, tg endpoints -d, tg beta clusters -d, tg beta clusters storage -d, tg beta jig secrets -d, tg beta jig volumes -d

Global parameters

The following parameters are available on every command:

Flag	Description
--help	Print help docs for the prefixed command.
--json	Forces the command into returning json response. Useful for scripting.
--non-interactive	Disables any prompting and manual input. In CI/Agent runners this is disabled by default.
--api-key [string]	Your Together API key. Defaults to reading environment variable TOGETHER_API_KEY.
--timeout [number]	Request timeout, defaults to 5 seconds.
--max-retries [number]	Maximum number of HTTP retries.
--version	Print CLI version.
--debug	Turns on debug mode to log diagnostic information.