CLI and MCP server: run Kite from a terminal or an AI agent

Kite ships two clients for the Platform API:

kite, a command-line tool for your terminal and scripts.
The Kite MCP server, which gives AI agents such as Claude Code, Claude, and Cursor the same actions as native tools.

Both call https://api.kiteml.com/v1 with your account, so anything you start from one shows up in the other and in the dashboard.

Install the CLI

The CLI is published on PyPI as kiteml-cli and needs Python 3.10 or later.

pip install "kiteml-cli[mcp]"

This installs two commands: kite, the CLI, and kite-mcp, the MCP server for running it locally. Drop [mcp] if you only want the CLI.

kite auth login

This opens the dashboard in your browser. Click Authorize CLI and the CLI receives an API key named CLI on <your machine>. It’s valid for 90 days and stored in ~/.kite/credentials.json. You can revoke it any time from Platform API → API keys. On a server or in CI, skip the browser and set an API key instead:

export KITE_API_KEY=kite_aBcDeFgH...

KITE_API_KEY takes precedence over a stored login. To check your setup, run kite doctor. It confirms the API is reachable and your key is valid, and exits non-zero if either check fails.

Command	What it does
`kite auth login`	Sign in through the browser. `--headless` prompts you to paste an API key instead.
`kite auth whoami`	Show the account you’re signed in as.
`kite auth logout`	Delete the stored key from this machine.
`kite doctor`	Check connectivity and authentication.
`kite config show`	Show the CLI’s settings, stored in `~/.kite/config.json`.

Use the CLI

Every command prints JSON: {"ok": true, "data": …} on success, or {"ok": false, "error": …} with exit code 1 on failure. Pipe it to jq in scripts. Commands that start long-running work take --wait to poll until the work finishes. Run kite --help or kite <command> --help to see every option.

Augmentations

kite augment create \
  --repo-id lerobot/pusht \
  -i "change the table surface to white marble, vary the lighting" \
  -n 20 \
  --wait

kite augment download aug_01J8X4... -o ./pusht-marble

download writes the dataset to disk with its LeRobot layout intact. To deliver to Hugging Face instead, pass --output huggingface --hf-repo your-org/pusht-marble. --model relight forces relighting, and --reference-image gives Kite a photo of the look to match. See Augmentations for when to use each.

Command	What it does
`kite augment create`	Start an augmentation.
`kite augment status <id>`	Status, progress and, once done, output files.
`kite augment list`	Your augmentations, newest first.
`kite augment cancel <id>`	Cancel a running augmentation. Unproduced episodes are refunded.
`kite augment download <id>`	Download a finished `download` augmentation.

Twins

Twins are in private beta and enabled per account. See Twins.

kite twin validate lerobot/svla_so101_pickplace
kite twin create lerobot/svla_so101_pickplace --out ./twins

A twin takes about 100 minutes. With --out, the command waits, then downloads the archive, verifies its checksum, and unpacks it to ./twins/<twin id>/scene.xml.

Command	What it does
`kite twin validate <dataset>`	Free check of a dataset. Creates nothing.
`kite twin create <dataset>`	Build a twin from one episode.
`kite twin status <id>`	Status, stage and progress.
`kite twin list`	Your twins, newest first.
`kite twin files <id>`	The files in a finished twin’s archive.
`kite twin cancel <id>`	Stop a twin. Finished stages stay cached.
`kite twin resume <id>`	Re-queue a failed or canceled twin from where it stopped.
`kite twin download <id>`	Download, verify, and unpack a finished twin.

RL runs

Describe the behavior in plain words, check the spec for free, then train:

kite rl plan open_duck_mini_v2 "walk forward at a steady pace" -o walk.json
kite rl validate walk.json
kite rl train walk.json --budget probe --wait
kite rl download rlr_01J8X4... -o .

plan writes a task spec you can edit. validate compiles it and reports what each reward term pays three canned policies, at no cost. --budget probe trains for 300 iterations to give a first signal in minutes. download unpacks the bundle to ./kiteml_<run id>/, with the policy at policy/policy.onnx. See RL runs for the spec and the bundle.

Command	What it does
`kite rl catalog`	Robots, objectives with their parameters, hardware, and budget presets.
`kite rl plan <robot> <prompt>`	A task spec from a description. Creates nothing.
`kite rl validate <spec>`	Free check of a spec in seconds.
`kite rl estimate <spec>`	Minutes and tokens per run.
`kite rl train <spec>`	Start one run per seed (`--seeds`, 1–5).
`kite rl status <id>`	Status, phase, progress and metrics.
`kite rl metrics <id>`	The training curve.
`kite rl report <id>`	The verdict and each check behind it.
`kite rl list`	Your runs, newest first.
`kite rl files <id>`	Every file in a packaged run’s bundle.
`kite rl download <id>`	Download and unpack the bundle.
`kite rl checkpoints <id>`	The iterations the run saved.
`kite rl checkpoint <id>`	Download one saved iteration.
`kite rl cancel <id>`	Stop a run. A run that started training keeps its policy so far.
`kite rl fork <id>`	Train from a run’s weights with the parameters you `--set`.
`kite rl continue <id>`	Keep training a run from where it stopped.

Connect the MCP server

The MCP server lets an AI agent drive Kite for you. Ask Claude to “relight lerobot/pusht to match this photo” or “train the Open Duck to walk and tell me when it passes”, and it calls Kite’s tools, polls the work, and reports back. Kite hosts the server at https://mcp.kiteml.com/mcp, so there’s nothing to install.

Claude Code
Claude
Cursor and others

claude mcp add --transport http kite https://mcp.kiteml.com/mcp

The first time Claude uses a Kite tool, your browser opens to sign in to Kite. To authenticate with an API key instead, for example in CI:

claude mcp add --transport http kite https://mcp.kiteml.com/mcp \
  --header "Authorization: Bearer $KITE_API_KEY"

In Claude or Claude Desktop, open Settings → Connectors, click Add custom connector, and enter:

https://mcp.kiteml.com/mcp

Claude opens a Kite sign-in window to connect your account.

Add the server to your client’s MCP configuration, for example ~/.cursor/mcp.json:

{
  "mcpServers": {
    "kite": {
      "url": "https://mcp.kiteml.com/mcp"
    }
  }
}

Clients that support MCP OAuth sign you in through the browser. Otherwise, add "headers": { "Authorization": "Bearer kite_..." } with your API key.

Run it locally

The hosted server can’t read or write files on your machine. When a twin or RL run finishes, its tools return the exact curl command to download the output, and your agent runs it. To give the server direct access to local files, for example to pass a reference photo by path, run it on your machine with the CLI installed:

claude mcp add kite -e KITE_API_KEY=$KITE_API_KEY -- kite-mcp

The local server reads KITE_API_KEY from its environment and runs over stdio.

Tools

Every tool returns {"ok": true, "data": …} or {"ok": false, "error": …}. Status tools for twins and RL runs add a next field that tells the agent what to do now: wait and check again, read the report, or run the download command.

Tools	What they do
`kite_augment_create`, `kite_augment_status`, `kite_augment_list`, `kite_augment_cancel`	Create and follow augmentations.
`kite_twin_validate`, `kite_twin_create`, `kite_twin_status`, `kite_twin_list`, `kite_twin_cancel`, `kite_twin_resume`	Check a dataset, then build and follow twins.
`kite_rl_catalog`, `kite_rl_plan`, `kite_rl_validate`	Explore what RL runs can train, and plan and check a spec for free.
`kite_rl_train`, `kite_rl_fork`, `kite_rl_cancel`	Start, branch from, and stop RL runs.
`kite_rl_status`, `kite_rl_metrics`, `kite_rl_report`, `kite_rl_list`	Follow RL runs and read their verdicts.
`kite_doctor`	Check connectivity and authentication.

Tools that only read are marked read-only, and tools that cancel work are marked destructive, so your client can ask before running them. Work the agent starts is charged to your account like any other request.

Need help? Reach out at raul@kiteml.com.

​Install the CLI

​Sign in

​Use the CLI

​Augmentations

​Twins

​RL runs

​Connect the MCP server

​Run it locally

​Tools

Install the CLI

Sign in

Use the CLI

Augmentations

Twins

RL runs

Connect the MCP server

Run it locally

Tools