Kite CLI and MCP server

Robot learning on cloud GPUs, from your terminal or your AI agent. Kite trains robot policies with reinforcement learning in simulation, augments LeRobot datasets, and rebuilds robot episodes as MuJoCo digital twins. This package gives you the kite command line and the kite-mcp server for Claude, Cursor, and any Model Context Protocol (MCP) client.

• RL training in simulation. Describe a behavior in plain English. Kite writes the task spec, checks its rewards for free, and trains the policy on a GPU. You get the policy as ONNX, the MuJoCo scene it trained in, and clips.
• Policy evaluation. Every RL run ends with a report: a pass, needs_review or fail verdict from measured checks (falls per minute, survival, command tracking, gait, posture), plus a vision model's read of the clip.
• Dataset augmentation. Relight a LeRobot dataset to match a deployment scene, or generate new variations of it with video augmentation. Output is a standard LeRobot dataset (Parquet + MP4), downloaded or pushed to Hugging Face.
• Digital twins. Point Kite at one episode of a LeRobot dataset and get an interactable MuJoCo scene of the room, with the objects the robot handles built to size.

Install

pip install "kiteml-cli[mcp]"

Python 3.10 or later. Drop [mcp] if you only want the CLI.

Sign in

kite auth login     # opens the dashboard; stores a 90-day API key
kite doctor         # checks the API is reachable and your key works

On a server or in CI, set an API key from app.kiteml.com instead:

export KITE_API_KEY=kite_...

Use it from an AI agent

Kite hosts the MCP server, so there's nothing to run:

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

Or run it locally over stdio with this package:

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

For Cursor and other clients, add {"mcpServers": {"kite": {"url": "https://mcp.kiteml.com/mcp"}}} to the MCP config. Then ask, for example: "train the Open Duck Mini to walk forward and tell me when it passes", or "relight lerobot/pusht to look like this photo".

Use it from the terminal

Train a walking policy for the Open Duck Mini v2:

kite rl plan open_duck_mini_v2 "walk forward at a steady pace" -o walk.json
kite rl validate walk.json                 # free: what each reward term pays canned policies
kite rl train walk.json --budget probe --wait
kite rl report rlr_...                     # the verdict and each check behind it
kite rl download rlr_...                   # -> kiteml_rlr_.../policy/policy.onnx

Augment a dataset:

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_... -o ./pusht-marble

Build a digital twin (private beta):

kite twin create lerobot/svla_so101_pickplace --out ./twins   # -> ./twins/twin_.../scene.xml

Every command prints JSON ({"ok": true, "data": ...}) so scripts and agents can parse it. Run kite --help for the full list.

Links

• Docs: CLI and MCP · RL runs · Augmentations · Twins
• Dashboard: app.kiteml.com
• Website: kiteml.com
• Support: raul@kiteml.com