> ## Documentation Index > Fetch the complete documentation index at: https://docs.cyberwave.com/llms.txt > Use this file to discover all available pages before exploring further. # Environment assistant > Describe a scene in plain language and the environment assistant builds it: robots, sensors, objects and, when the catalog has none, new assets. The **environment assistant** builds and edits the world your robots live in. You describe the scene, for example *"a workbench with an SO-101, a camera above it and two bins"*, and it finds the right twins in the catalog, places them and proposes the changes for you to accept. When nothing in the catalog fits, it can generate a new asset. ## What it does | You ask for | The assistant | | ---------------------------------- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------- | | A scene or a change to it | Searches the catalog first, places twins, and uses simple procedural primitives for basic geometry | | An object the catalog doesn't have | Generates a **private URDF asset** in the background while you keep editing ([Generate assets](/feature-reference/environment-editor/assistant-asset-generation)) | | A change to a generated asset | Builds a new candidate revision; the current asset stays unchanged until you accept the preview | | Help with a reference | Takes tagged scene objects and reference images alongside your message | Generated assets start private and scoped to the environment that asked for them. Sharing them further follows the normal asset visibility controls. ## Use it in the editor 1. Open an environment and select **Agent** in the right-hand panel. 2. Describe what you want. Tag scene objects or attach a reference image if it helps. 3. Review what the assistant proposes, then accept it. The conversation stays the same when you switch between **Edit** and **Monitor** in Simulate or Live, within the same page session. Unsent drafts, tagged objects and reference images are kept when you visit **Control** and come back. In **Monitor**, action cards are read-only, and switching to Monitor asks the assistant to stop unfinished edits without stopping running robots or training jobs. To act on the robots themselves, use the [control agent](/ai/control-agent); it has its own composer. More on the editor: [Environment editor](/feature-reference/environment-editor). ## Use it from Python ```python theme={null} from cyberwave import Cyberwave cw = Cyberwave() # Create a new environment from a description env = cw.agents.environment.create("A workbench with an SO-101 arm, a camera above it and two bins") # Continue the conversation in an existing environment reply = cw.agents.environment.message(environment_uuid, "Add a second camera on the left side") # Models that can run the assistant print(cw.agents.environment.models()) ``` `message()` also accepts the conversation `history`, a `pending_confirmation` to answer, and an image (`image_base64`, `image_mime_type`) as a reference. ## From your AI tools (MCP) With the [MCP server](/overview/tools/mcp-server), Claude Code, Cursor or any MCP client can plan a scene with `cw_plan_scene` and create environments with `cw_create_environment`. `cw_plan_scene` is advisory: it returns a plan and checks, and changes nothing. ## Next steps Once the scene is ready, tell the robots what to do. Build stages, revisions and limits for generated URDF assets.