AHa-3D, titled "Agentic Tool Use for Real2Sim with GPT-6 Astra", turns an ordinary indoor video into an editable 3D scene in Blender. It rebuilds the room from reference geometry and a library of reusable furniture and materials, estimates the motion of the people in the video, places them in the shared room with foot-ground and object contact refinement, and renders the result with the original camera and timing. It can also generate new actions for people in the scene and export saved scenes as interactive browser demos. The README's demo is an office rebuilt from a 10-second video and rendered through the recovered source camera, with an editable scene in the examples folder that opens with Blender 5.2. A single command, tools/indoor, drives scene work from request to delivery with resumable runs, each in its own directory, and subcommands for intake, preflight, batch queuing, preview rendering, placement checks and acceptance tracking. The asset library covers furniture, plants, articulated cabinets and procedural materials, and saved scenes can have furniture and materials swapped while keeping IDs and orientation. The installer asks you to choose one of three levels before changing an environment: Static (editable room), Human reconstruction (adds PMPose, GVHMR and Blender body skinning) or Robotics motion generation (adds Kimodo and native G1 motion). It needs one Linux workstation (Ubuntu or WSL2) with an NVIDIA GPU, Python 3.11 and a compatible Blender; human skinning also needs Blender 4.5 and its body extension. The runtime policy targets a single 24 GB GPU but was originally validated only on one 96 GB RTX PRO 6000, and smaller cards are untested. Model checkpoints (such as Pi3X and SAM3), SMPL-X and other body assets, and Blender are obtained separately under their own terms and are not bundled. The README adds that Opus 5.5 can also drive the repository based on initial experiments but has not been fully tested. Project-authored code, skills and assets are under Apache-2.0.