# AI use cases for Agentic STAR (makezur/agenticSTAR) Repo: [makezur/agenticSTAR](https://github.com/makezur/agenticSTAR) | Use cases: 3 | Page: https://dedupedai.com/use-cases/makezur-agenticstar/ AgentSTAR is a harness where a coding agent (Codex or Claude Code) reads a monocular video, builds an articulated object in Blender, poses it per frame, and iterates until the renders match the frames. This file holds every use case of this repo. Each one starts with a "## " heading and ends with its URL. - Reconstruct the garden shears example end to end: https://dedupedai.com/use-cases/makezur-agenticstar/run-garden-shears-example/ - Turn your own video of an object into a 3D capture: https://dedupedai.com/use-cases/makezur-agenticstar/reconstruct-own-video-object/ - Compare Codex and Claude Code on the same capture: https://dedupedai.com/use-cases/makezur-agenticstar/compare-codex-and-claude-runs/ ## How to reconstruct the garden shears example end to end with Agentic STAR Repo: [makezur/agenticSTAR](https://github.com/makezur/agenticSTAR) | Difficulty: Advanced | Time: 5 hours | Basis: From the README Run one supervised agent session on the committed example and get a GLB and pose.json. **Who it is for:** A researcher or engineer who wants to see the full pipeline work on a known capture before using their own video. **The problem:** Setting up an agentic reconstruction harness is hard to judge without seeing a complete run on known data. **The outcome:** A runs/ directory containing scene.py, mesh/object.glb, mesh/pose.json and per-pass renders under iterations/. ### Steps 1. Run `./install.sh` to build the artscript env, Blender 4.2.5 and the pi3x env. 2. Run `export ANTHROPIC_API_KEY=sk-ant-...` in the shell you will use. 3. Run `tools/install_claude_bwrap.sh` once so the key is stored in a state dir outside the checkout. 4. Run `tools/run_kf.sh --agent claude --timeout-hours 4 examples/garden_shears/capture:0`. 5. Run `tools/claude-supervisor attach --run-dir runs/capture_kf_` to watch the run. ### Prompt to paste into your agent ```text In the agenticSTAR repo, set up the environment with ./install.sh and run the garden shears example with Claude Code using tools/run_kf.sh on examples/garden_shears/capture:0 on GPU 0. Expect runs/ to contain scene.py, mesh/object.glb and mesh/pose.json. Verify that mesh/pose.json holds per-frame poses and joint states before reporting success. ``` ### Caveat A run takes hours and many tokens, so budget for model cost and use --timeout-hours to bound it. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/makezur-agenticstar/run-garden-shears-example/ ## How to turn your own video of an object into a 3D capture with Agentic STAR Repo: [makezur/agenticSTAR](https://github.com/makezur/agenticSTAR) | Difficulty: Advanced | Time: 6 hours | Basis: From the README Extract frames, add object masks, run Pi3X cameras, then let the agent rebuild the object. **Who it is for:** A robotics or vision researcher with a phone video of a hinged or articulated object who needs a joint-level model. **The problem:** Monocular video gives camera motion and appearance but not a posed, jointed 3D model that can be inspected or reused. **The outcome:** A capture directory and an agent run that produces mesh/object.glb with named parts and mesh/pose.json with per-frame joint states. ### Steps 1. Run `tools/video_to_frames.sh clip.mov my_input --every 5`. 2. Add binary object masks as my_input/mask_object/ .png for those frames, and optionally mask_hand/ .png. 3. Run `micromamba run -n pi3x python tools/make_capture.py --src my_input --out captures/my_object`. 4. Run `tools/run_kf.sh --agent claude captures/my_object:0`. 5. Check the output under runs/ for mesh/object.glb and mesh/pose.json. ### Prompt to paste into your agent ```text Using the agenticSTAR repo, convert into a capture with tools/video_to_frames.sh and tools/make_capture.py, then run tools/run_kf.sh with --agent claude on the resulting capture. Expect mesh/object.glb and mesh/pose.json under runs/. Verify that the mask folder has one PNG per extracted frame before starting the run. ``` ### Caveat Masks can come from any segmenter, but you must supply them yourself, and the README does not describe how to produce them. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/makezur-agenticstar/reconstruct-own-video-object/ ## How to compare Codex and Claude Code on the same capture with Agentic STAR Repo: [makezur/agenticSTAR](https://github.com/makezur/agenticSTAR) | Difficulty: Advanced | Time: 10 hours | Basis: From the README Run the same capture with both supported agents and compare the resulting poses. **Who it is for:** A researcher evaluating which coding agent produces better articulated reconstructions on a fixed capture. **The problem:** Agent choice changes results, and the repo supports two agents with separate supervisors. **The outcome:** Two runs on one capture, each with its own mesh/pose.json and iteration renders that can be compared side by side. ### Steps 1. Run `export OPENAI_API_KEY=sk-...` and then `tools/install_codex_bwrap.sh` once. 2. Run `tools/run_kf.sh --agent codex examples/garden_shears/capture:0`. (inferred, check the README) 3. Run `tools/agent-supervisor` to watch the Codex run. 4. Run `tools/run_kf.sh --agent claude --timeout-hours 4 examples/garden_shears/capture:0` for the Claude run. 5. Compare the mesh/pose.json and side-by-sides under iterations/ from each runs/ directory. ### Prompt to paste into your agent ```text In the agenticSTAR repo, run tools/run_kf.sh on examples/garden_shears/capture:0 once with --agent codex and once with --agent claude. Expect two runs/ directories, each with mesh/pose.json. Verify that both runs finished or hit their timeout before comparing outputs. ``` ### Caveat The paper results used a critic branch with GPT-5.6-Sol, and the README says newer models should work without it, so a fair comparison needs the same settings on both sides. Risk, credentials: Both API keys are needed, so keep them out of shared logs and terminal history, since the install scripts store the Claude key in a state dir outside the checkout. Disclaimer: Third-party repos we don't own or vouch for. AI-generated use cases can be wrong: review everything and run at your own risk. Full text: https://dedupedai.com/disclaimer/ URL: https://dedupedai.com/use-cases/makezur-agenticstar/compare-codex-and-claude-runs/