SpatialClaw is a research evaluation harness for training-free spatial reasoning, in which a VLM agent uses Python code as its action interface: it plans, writes one code cell per step, passes an AST safety check, executes in a persistent Jupyter kernel pre-loaded with perception primitives (SAM3 segmentation, Depth-Anything-3/Pi3/MapAnything 3D reconstruction, EasyOCR) and scientific libraries, then receives condensed stdout/variable/image feedback until it emits a ReturnAnswer sentinel. The loop is assembled as a LangGraph state machine (init -> plan -> llm_step -> execute -> feedback -> reflection -> terminate) with code, ReAct, and single-pass executor variants plus a no-tool CoT baseline. This is the system that generated 14k+ executable video-QA demonstrations for NVIDIA Nemotron 3.5 Super VL per docs/nemotron.md. The runtime is a three-service distributed architecture: vLLM OpenAI-compatible LLM servers, a FastAPI GPU tool server exposing segmentation/reconstruction/OCR over a type-safe MessagePack RPC codec, and CPU-side agent processes managing Jupyter kernel pools. Services discover each other via file-locked JSON registries (serve.json, gpu_server.json) and are orchestrated as auto-resuming, overlapping SLURM chain jobs with sendmail-like rotate-and-restart behavior. It ships interactive CLIs for vllm/gpu-server/agent lifecycle management, a FastAPI results visualization server, and a GPU monitoring dashboard. Key entry points are python -m spatial_agent.entrypoints.run (benchmark evaluation), cot_baseline.py, launch_vllm.py, and launch_gpu_server.py. On the data side, 20 benchmark loaders (BLINK, ERQA, Omni3D, MMSI, VSI-Bench, Video-MME-v2, SPAR-Bench, CV-Bench, etc.) are registered through a factory pattern with per-benchmark exact-match and Mean Relative Accuracy (MRA) threshold scoring, category/question-type aggregation, deterministic subsampling, JSONL prediction resume, and per-sample trace/HTML reports. Datasets are fetched externally into data/ by scripts/download_datasets.sh; third-party vision models are pinned git submodules. This is application-class research infrastructure rather than a reusable library or data pipeline. It is tightly coupled to NVIDIA-internal SLURM accounts/partitions, conda envs, nightly vLLM wheels (CUDA 12.9), HF offline caches, and gated weights; it will not run outside the original cluster without significant reconfiguration. No test suite is present, several modules carry acknowledged hacks (a placeholder crc32 in the hand-rolled TFRecord parser, regex-only security sandbox, GPU idle-reaper keepalives), and chain managers/benchmark evals show duplication. The genuinely reusable assets are the RPC codec, MRA scoring utilities, lazy video loading scaffolding, aligned per-frame data model, and LLM error-condensation logic.