Agentic quickstart

Choose how you want to start: prompt Cyberwave’s built-in agent in your browser, or connect an agent you already use.

Prompt in Cyberwave
Connect your agent

Build with the built-in agent. Describe your scene, review what it creates, and refine it with follow-up prompts. You need a Cyberwave account and a browser; no API key or coding-agent installation is required.

Watch the browser walkthrough

This 2-minute 20-second walkthrough shows a fresh scene built from the dashboard and a separate follow-up edit. Captions describe the silent footage. Build waits are shortened; robot actuation and rendered camera capture are next steps, not results demonstrated in this video.

Walkthrough transcript

0:00 — Start with a prompt. Build a robotics scene using Cyberwave’s built-in agent. The browser path needs no external coding-agent installation.0:10 — Describe your scene. Select your workspace on the dashboard. Ask for a Go2, camera, two orange boxes and a floor, then select Build environment.0:25 — Watch the build. The agent finds assets, applies scene changes and checks its result. The recording accelerates this sequence. Asset choices may change during validation.1:05 — Review the result. The completed scene contains Unitree Go2, Standard Camera, two orange boxes and a floor. This is a scene-editing result.1:20 — Refine conversationally. A separate earlier take shows a follow-up prompt adding a floor while preserving existing objects.1:40 — Check readiness. Inspect controls and simulation prerequisites before running. The shown camera still needs setup; this video does not demonstrate verified motion or rendered camera capture.2:05 — Keep building. Continue below, or use the optional external-agent and Python paths.

1. Open your workspace

Create a Cyberwave account, finish verification, and create or join a workspace. On the dashboard, find What do you want to build? and select the workspace for your scene. You can start directly from this prompt box; creating a blank environment first is optional.

2. Describe your scene

Paste this into Describe the environment you want to create, then select Build environment:

Build a simple inspection scene with a Unitree Go2, an RGB camera facing it,
two orange boxes beside the robot, and a flat floor beneath them.
Leave a clear path in front of the Go2.
Use catalog assets and simple primitives. Do not start simulation or connect
physical hardware yet.

The editor opens while the assistant builds. If the editor tour appears, choose Later to follow this walkthrough, or take the tour first. Watch the assistant find assets and apply scene changes. Review any proposed changes or confirmation cards before accepting them. Check the scene itself: you should see one Go2, one camera, two orange boxes and a floor. Wait for the final build summary; the agent may revise its asset choices during its checks.Try a follow-up: in Edit mode, open the right-hand Assistant panel. Its heading should say Environment assistant. If hidden, use Panels to show it. The agent can keep editing the same scene.

Add a flat floor beneath the Go2 and boxes if one is missing.
Keep the clear forward path and existing objects. Use a simple primitive.
Make scene edits only; do not start simulation or preparation tests.

Already have a Go2 and camera? Use this continuation instead of duplicating them:

Add two orange boxes beside the Go2 to make a simple inspection scene.
Keep a clear path in front of the robot and position the existing RGB camera
to see it. Reuse the existing robot and camera. Make scene edits only;
do not start simulation, preparation tests, training, asset generation,
or connect hardware.

Watch: create and refine with prompts

These silent review clips show the continuation prompts above in an existing Go2 scene. They use actual browser captures at 8 fps with pauses shortened.

The clips demonstrate scene editing; simulation movement and camera capture are separate steps below.

Create something beyond the catalog

Describe the object, its dimensions and how it should behave, or attach a reference image. The environment assistant searches existing assets first; supported new objects can be generated as private assets. Review the build status and preview before accepting a revision. Generation and simulation preparation are separate operations—check their requirements before starting. Explore asset creation.

3. Ask the robot to act

Scene editing places objects. To make the robot act, switch to Simulate and open the Control assistant. Ask it to inspect readiness first:

Inspect this Go2's simulation readiness and available controls.
Explain anything missing and which simulation engine this task needs.
Do not start compute or execute movement yet.

MuJoCo starts billable cloud compute and needs credits. The free Playground supports kinematic motion but does not render camera frames. Check your workspace balance and review preparation requirements before starting MuJoCo. Stop it when you finish.

When readiness checks pass, start a bounded MuJoCo session in the simulation controls. Then ask:

In simulation, plan a short forward movement for the Go2, followed by a stop.
Use its supported bounded controls and show me the plan before execution.

Review the target robot, direction and limits, then approve the supported action. Watch the robot and inspect the reported result. Follow up with:

Check the robot's actual pose before and after the action. Tell me what
movement was observed and whether it stopped. If motion cannot be verified,
explain what failed.

Prefer an arm? Build the scene with an SO-101, then ask for a small movement of one controllable joint within its limits. Review and approve that action.

4. Inspect the camera view

Check the RGB camera’s simulation readiness before starting the camera test. If the viewer reports needs simulation setup, open Fix simulation and review Prepare simulation. That preparation can use credits. If the camera was added after startup, restart MuJoCo to load the updated scene.Ask the assistant:

Inspect the RGB camera in this simulation. Capture a fresh rendered frame
if supported and describe what is visible. Report any missing setup or
unavailable capability. Do not substitute an editor preview for a camera frame.

Open the returned frame and check that it shows the scene. If the assistant cannot retrieve a frame, switch to Connect your agent for the SDK camera example, or follow the reported readiness issue; a scene preview alone does not complete this step.

5. Stop and keep building

Stop the simulation in the environment controls and confirm No active simulation. Your environment remains available for further editing.Return to Edit → Assistant to change the scene or ask for a workflow draft:

Draft an inspection workflow for this scene: inspect the two boxes, capture
an image when the camera is ready, then stop. Explain any missing capabilities.
Do not run the workflow yet.

Review the draft and its requirements before running it. Learn more about the environment assistant, control agent and workflow assistant.

Use Claude Code, Codex, another MCP client or Python. Connect your workspace, then create and control a simulated robot from your own tools.You’ll need Git and an installed coding agent for the skill setup. Python examples need Python 3.11+. Terminal commands support macOS, Linux and Windows WSL.

1. Set up your workspace

Create a Cyberwave account, finish verification, and create or join a workspace. Keep Cyberwave open beside your terminal to watch the scene.

The motion and camera examples use MuJoCo, which starts billable cloud compute and needs credits. Check your workspace balance before starting. The free Playground supports kinematic motion but does not render camera frames.

2. Get your API key

Open Profile → Access, create an API key named Quickstart, and scope it to one workspace. Copy it when shown. See API keys and scopes for details.In your terminal, run this block and paste the key at the hidden prompt. It works in Bash and Zsh, including WSL, and keeps the value out of your command history:

printf 'Cyberwave API key: '
read -rs CYBERWAVE_API_KEY
printf '\n'
export CYBERWAVE_API_KEY

Stuck? API key setup

The key is shown once; create a replacement if you missed it. For workspace_required, create a key scoped to one workspace. For 403, check your workspace role and key scope. Keep the key in your terminal or client’s secret settings. Ask for help on Discord.

3. Choose your client

Claude Code
Codex
Other MCP clients
Python SDK

Install the skill and register the hosted MCP server:

git clone https://github.com/cyberwave-os/cyberwave-skill ~/.claude/skills/cyberwave
claude mcp add --scope user --transport http cyberwave https://mcp.cyberwave.com/mcp --header "Authorization: Bearer $CYBERWAVE_API_KEY"
claude

Claude stores the expanded authentication header in its local user configuration. In Claude Code, invoke /cyberwave. Claude Code MCP setup describes connection status and configuration scopes.

Install the skill and register the hosted server:

git clone https://github.com/cyberwave-os/cyberwave-skill ~/.agents/skills/cyberwave
codex mcp add cyberwave --url https://mcp.cyberwave.com/mcp --bearer-token-env-var CYBERWAVE_API_KEY
codex

Start Codex from the terminal where you exported the key; the server reads that environment variable when connecting. Invoke $cyberwave. See Codex MCP setup.

In your client’s MCP settings, add a Streamable HTTP server named cyberwave at https://mcp.cyberwave.com/mcp. Configure its Authorization header as Bearer followed by your API key using the client’s private credential field. Reconnect the server.Header and secret support varies by client. Use the MCP client guide for configuration details; if your client cannot send a Bearer header, use Claude Code, Codex or the SDK tab. Interactive cyberwave login does not configure a hosted MCP client.

In the terminal where you exported your key:

pip install cyberwave

Use the Python motion example. It creates the environment automatically and prints a link to open it. The SDK reference covers installation extras and advanced usage.

Verify your MCP connection: if you chose Claude Code, Codex or another MCP client, ask your agent:

List my Cyberwave workspaces using the Cyberwave MCP tools.

A tool response listing your workspace confirms the connection. If there are several, select it by name. Continue with the same agent session. For Python, continue with the SDK motion example below.

Stuck? Connection failed

For 401 or AUTH_ERROR, replace the key and update your client’s credential. If tools are missing, reconnect the MCP server and restart the agent. If the skill folder already exists, reuse the installed skill instead of cloning over it. Ask for help on Discord.

4. Move a simulated robot

Go2
Prefer an arm?

Copy these prompts into your agent, one at a time:

Create a new Quickstart Environment in my workspace with a Unitree Go2.
Use simulation only. Start a MuJoCo simulation for five minutes, inspect
its readiness and available controls, and give me the environment link.
Discover resource references yourself and reuse them in later calls.

Open that link, then ask:

In that simulation, walk the Go2 forward 1 metre, turn left 90 degrees,
then stop. Plan and dispatch supported bounded controls. Inspect actual
pose telemetry before and after; report observed motion separately from
commands accepted. Do not substitute scene editing for robot actuation.

Render a preview of the environment and tell me where the robot ended up.
If movement cannot be verified, explain the failed step instead of
claiming the robot moved.

SDK alternative: save as quickstart.py, then run python quickstart.py:

import time
from cyberwave import Cyberwave

cw = Cyberwave()
robot = cw.twin("unitree/go2")
cw.affect("simulation", duration=300)
robot.policy.ensure_attached()
time.sleep(2)
robot.move_forward(0.5, duration=2)
cw.disconnect()

move_forward takes speed in metres per second, then sends a stop at the end of the duration. This requests roughly one metre of travel; inspect the robot in the environment to verify motion. Disconnecting the SDK leaves the simulation running until its duration expires or you stop it.

Use the same setup, then ask:

Create a new Quickstart Environment with the-robot-studio/so101 and start
a five-minute MuJoCo simulation. Discover resource references yourself.
Inspect the arm's controllable joints and limits. Plan and execute a small
bounded movement of one joint in simulation, then run a saved motion pose
if one is available. Verify actual joint telemetry and render a preview.
Give me the environment link. Do not substitute editor transforms for motion.

SDK alternative:

import time
from cyberwave import Cyberwave

cw = Cyberwave()
robot = cw.twin("the-robot-studio/so101")
cw.affect("simulation", duration=300)
robot.policy.ensure_attached()
time.sleep(2)
robot.joints.set("shoulder_pan", 15, degrees=True)
cw.disconnect()

Stuck? The robot isn't moving

If simulation startup fails, check credits and its status in the environment. If the controller is still attaching, wait for readiness before retrying a bounded command. If a command is accepted without observed movement, ask your agent to inspect simulation control readiness and telemetry; an accepted command alone does not prove motion. Ask for help on Discord.

5. Capture a camera frame

Continue in the same agent session:

In the same simulated environment, add an RGB camera, positioned to see the
robot. Check that the camera supports rendered frames and is ready for MuJoCo.
If simulation setup is missing, explain the required preparation before
starting it. Inspect whether the active MuJoCo scene includes the new camera; if
needed, stop and restart the simulation for five minutes to load it. Capture
one fresh camera frame and describe what is visible. Report the frame's
freshness; an environment preview is not a camera frame.

For the SDK path, create a camera before cw.affect(...) in your motion example script:

camera = cw.twin("cyberwave/standard-cam")
camera.edit_position(x=2.0, y=0.0, z=1.5)

After simulation starts, capture a frame before disconnecting:

frame = camera.get_frame("path")
print("Camera frame:", frame)

Open the saved image to inspect it. Ask your agent to describe the image if it supports images. When finished, stop the simulation in the environment; ask your agent to stop it and verify its status if you used MCP.

Stuck? Camera capture failed

SimulationNotRunningError means a running MuJoCo simulation is needed. If the viewer says the camera needs simulation setup, open Fix simulation for that camera and review Prepare simulation. Preparation includes a simulator test and may use credits; complete it before restarting MuJoCo. If the camera was added after startup, restart the simulation to load the updated scene. If the frame is missing or stale, inspect camera and simulation readiness before capturing again. Ask for help on Discord.

6. Stop the simulation

Stop the simulation in the environment controls and confirm No active simulation. Closing your terminal or calling cw.disconnect() only disconnects the client; it does not stop the simulation.

Next steps

Fork a template

Explore existing templates and clone one into your workspace.

Build with the community

Share your simulation and find your next project.

Run on real hardware

Pair hardware when you’re ready to switch from simulation to live control.

Get started

Build with AI

Simulate & test

Connect hardware

Train & improve

Deploy & operate

Watch the browser walkthrough

1. Open your workspace

2. Describe your scene

Watch: create and refine with prompts

3. Ask the robot to act

4. Inspect the camera view

5. Stop and keep building

1. Set up your workspace

2. Get your API key

3. Choose your client

4. Move a simulated robot

5. Capture a camera frame

6. Stop the simulation

Next steps

Fork a template

Build with the community

Run on real hardware

​Watch the browser walkthrough

​1. Open your workspace

​2. Describe your scene

​Watch: create and refine with prompts

​3. Ask the robot to act

​4. Inspect the camera view

​5. Stop and keep building

​1. Set up your workspace

​2. Get your API key

​3. Choose your client

​4. Move a simulated robot

​5. Capture a camera frame

​6. Stop the simulation

​Next steps

Fork a template

Build with the community

Run on real hardware

Watch the browser walkthrough

1. Open your workspace

2. Describe your scene

Watch: create and refine with prompts

3. Ask the robot to act

4. Inspect the camera view

5. Stop and keep building

1. Set up your workspace

2. Get your API key

3. Choose your client

4. Move a simulated robot

5. Capture a camera frame

6. Stop the simulation

Next steps