Each prim shows: path, type, world position. For large scenes, use dump_scene
    which also includes properties and bounding boxes.
    

    Types: Cube, Sphere, Cylinder, Cone, Capsule, Xform, Camera, DistantLight, SphereLight, etc.
    Or load a USD file via usd_path. A prim already at prim_path keeps its type: one of another type is refused.
    

    1. **Viewport image** (PNG) -- what the camera sees
    2. **Bounding boxes** (TXT) -- screen-space bounding boxes mapping each visible prim to pixel coordinates
    3. **Instance segmentation** (PNG) -- each prim rendered as a unique color, with a legend mapping colors to prim paths

    Use files 2 and 3 to understand which prim is where in the viewport image.
    Set include_image=True to return the viewport image inline (uses more context but lets you see the scene directly).
    

    Auto-fits camera distance so the prim fills the viewport with ~10px margin.
    Captures from all 6 face centers + 8 cube corners by default, giving full coverage.
    Saves each angle as a PNG file. Use Read tool to view images.
    

    Actions:
    - "get": Return current lighting mode, available modes, and scene lights
    - "set": Set lighting mode. Available modes: "off", "camera", "stage",
      "Default", "Colored Lights", "Grey Studio" (light rigs)
    - "set_camera_light": legacy, turns the camera light on or off

    When capture_viewport returns a black image, use this tool to check and change
    the lighting mode, then recapture.
    

    Available: franka, ur10, carter, jetbot, g1, go1, go2, h1, spot, anymal. Or a USD file as robot_type: a path
    on the Kit app's machine, a URL, or a path in Isaac Sim's asset library (/Isaac/...). Without prim_path,
    the prim is named after the file.
    

    Radians for revolute joints and metres for prismatic ones, velocities per second (set_joint_targets takes
    degrees). Needs Isaac Sim; tested on 6.2.
    

    Isaac Sim draws with isaacsim.util.debug_draw, turned on at the first call. Its clear takes every line drawn
    with it, so when a drawing expires, lines other code drew with it go too. Other Kit apps draw with
    omni.debugdraw when they have it. A new stage drops the old one's drawings.

    Records the active viewport, from its camera at its own size: aim it with set_camera or look_at_prim first.
    camera_path, width and height are ignored for now (#25); the answer warns when given a camera_path, or a size
    other than 640x480.

    Captures a frame every few app updates, about 1/fps seconds, whether or not the simulation plays: play it to
    record motion. Frames are saved as PNGs in mcp_output/recordings/. Use stop_recording when done,
    then get_recording_frame to review any frame.

    Optionally tracks prim state (transforms, properties) alongside video frames.
    State is written to a grep-friendly state.txt in prim-block format.
    

    Returns session info and frame count. Use get_recording_frame to review frames.
    

    Like get_scene_tree but focused on geometry budget: each mesh shows
    face/vertex/triangle counts, each group shows subtree_faces total.
    

    If input_path is omitted, flattens the current stage.
    

    Two modes:
    1. Prim comparison: provide prim_path_a and prim_path_b
    2. Variant comparison: provide prim_path, variant_set, variant_a, variant_b

    Returns side-by-side stats with deltas. Properly clears BBoxCache between
    variant switches to avoid stale data.
    

    Updates both relationship targets (material bindings) and asset path
    attributes (texture file paths) that start with old_prefix.
    

    The script has access to: omni, carb, Usd, UsdGeom, UsdLux, UsdPhysics, Sdf, Gf, mcp.
    Assign to 'result' to return data. stdout/stderr are captured.

    The `mcp` object provides async access to MCP handlers from within scripts.
    Use `await` to call its methods (the script automatically runs as async):

        viewport = await mcp.capture_viewport(width=1280)
        tree = await mcp.scene_tree()
        await mcp.set_camera([5, 3, 5], target=[0, 0, 0])
        result = {"tree": tree["result"]}

    Available mcp methods:
        mcp.capture_viewport(width, height, camera_path)
        mcp.set_camera(position, target)
        mcp.look_at(prim_path, distance, azimuth, elevation)
        mcp.inspect(prim_path, angles, width, height, zoom_offset=0.0, isolate=False)
        mcp.scene_tree(root, max_depth)
        mcp.prim_properties(prim_path)
        mcp.prim_bounds(prim_path)
        mcp.set_transform(prim_path, position, rotation, scale)
        mcp.create_prim(prim_path, prim_type, **kwargs)
        mcp.delete_prim(prim_path)
        mcp.set_material(prim_path, color, opacity, roughness, metallic)
        mcp.sim_control(action)  -- play/pause/stop/step
        mcp.sim_state()
        mcp.get_logs(count, min_level, channel, search)

    Each returns the full handler response dict with "status" and "result" keys.

    Use Context7 MCP (library: isaac-sim/isaacsim) to look up the correct API calls.
    

    Use this to investigate errors: when any tool fails, call get_logs() to see what the Kit app reported. Kit's
    file holds Info and above by default; verbose lines need Kit started with both --/log/level=verbose and
    --/log/fileLogLevel=verbose.