NeuroClaw: Reproducible Neuroimaging Pipelines

• NeuroClaw is a domain-specialized multi-agent research assistant that unifies raw neuroimaging workflows into auditable and reproducible pipelines.
• Its three-tier architecture integrates user intent parsing, DAG planning, and verified tool execution to handle sMRI, fMRI, dMRI, and EEG data.
• By enforcing BIDS compliance and systematic audit checkpointing, it enhances reliability, reduces manual curation, and benchmarks reproducibility.

NeuroClaw is a domain-specialized multi-agent research assistant designed to make neuroimaging workflows executable and reproducible from truly raw data, integrating agent systems, harness engineering, and reproducibility benchmarking. Built to operate across neuroimaging modalities such as sMRI, fMRI, dMRI, and EEG, NeuroClaw addresses persistent challenges in the field—heterogeneous data formats, complex multi-stage toolchains, fragility in environment specification, and nontrivial reproducibility gaps—by orchestrating transparent, semantically grounded, and audit-traceable pipelines without manual curation or bespoke coding (Wang et al., 27 Apr 2026).

1. Motivation and Problem Scope

Neuroimaging research is characterized by its dependency on diverse data types (DICOM, NIfTI, EEG), multistep pipelines, and the aggregation of numerous manual conventions and third-party tools. Routine workflows encompass steps such as DICOM-to-NIfTI conversion, BIDS reorganization, skull stripping, nonlinear registration (e.g., ANTs), FreeSurfer surface reconstruction, denoising (e.g., fMRIPrep), diffusion/tractography, and quality control. Each stage imposes unique and evolving metadata, computation, and validation demands. Even minor metadata misalignments, environment drifts, or version inconsistencies can cascade into analysis failure, rendering reproducibility a chronic practical obstacle. The field’s prevalent reliance on unpinned Docker images, fragile conda environments, and insufficiently audited scripts impedes reliably replayable workflows across laboratories or computing platforms.

NeuroClaw emerged to address these issues fundamentally, seeking to unify how raw neuroimaging data can be processed, how tool chains are executed and audited, and how every analysis step is tracked, rerunnable, and verifiable irrespective of local computing circumstances (Wang et al., 27 Apr 2026).

2. System Architecture: Three-Tier Skill/Agent Hierarchy

At its core, NeuroClaw adopts a three-layer hierarchical architecture that decomposes complex research workflows into modular, auditable, and reusable “skills”:

1. Interface Layer (User-Facing): Parses free-form user intents and translates them into research goals. Requests are routed to appropriate subagents, and user-facing reports and summaries are generated and returned.
2. Subagent (Orchestration) Layer: Maintains a directed acyclic graph (DAG) that encodes workflow stage dependencies. Decomposes user-specified goals into specific modality, tool, dataset, and model skills. Ensures the correct, ordered invocation of underlying operations, applying explicit pre- and post-condition logic at each edge.
3. Base Skill (Tool-Invocation) Layer: Encapsulates atomic, verifiable operations such as file format conversion, metadata validation, and execution of processing commands (e.g., antsRegistration, fsl_anat, recon-all). Each base skill executes in a pinned, environment-managed context, performs checkpoint/snapshotting before execution, verifies artifact existence and integrity after execution (including checksumming and NaN/∞-scan), and emits structured audit logs in machine-readable form.

The procedural model realizing these layers discards monolithic, opaque prompts in favor of atomic, self-verifying skill invocations. This ensures modularity, transparency, and direct traceability of each pipeline stage:

goal_spec = InterfaceAgent.parse_intent(user_query) skill_dag = OrchestrationAgent.plan(goal_spec) for skill in skill_dag.topological_sort(): env = Harness.setup_environment(skill.requires) result = BaseSkill.execute(skill.command, env) if not BaseSkill.verify(result, skill.checkpoints): raise ExecutionError(f"Verification failed at {skill.name}") Harness.checkpoint_state(skill, result) report = InterfaceAgent.compose_report(skill_dag.audit_logs) return report

This approach reduces execution drift, confusion, and token bloat, while enforcing rigorous pre/post-conditions for each processing unit.

3. Data Handling and Environment Control

NeuroClaw operates on genuinely raw neuroimaging data, including uncurated DICOM directories and NIfTI archives from any vendor. Header fields are inspected to infer modal identity, and data are reorganized into BIDS-compliant layouts without requiring user-side pre-processing. BIDS (Brain Imaging Data Structure) compliance underpins all Subagent logic, ensuring consistent handling of metadata, naming, and modality distinctions (e.g., differentiating HCP from ADNI conventions and mapping echo times or gradient tables as required).

The environment management system is fully specified, eliminating reproducibility gaps due to version drift or implicit state:

Component	Method	Verification Mechanism
Python Environments	Pinned YAML specifications	Version manifests, hash checks
Tool Installers	Automated shell scripts	Executable checks, version assertion
Docker	Image hash, checksum option	Container integrity scan, isolation
GPU Config	Auto-detection, driver match	Runtime-check, manifest logging

Every executed command logs environment specifications, timestamps, input/output file checksums, and QC flags (e.g., missing files, MRI QC metrics) to a manifest, allowing subsequent rerun, state replay, or focused debugging.

4. Auditability and Checkpointing

One of NeuroClaw’s defining characteristics is the systematic structuring of audit traces at each pipeline step. Each Base Skill invocation:

• Snapshots and records the current input state before execution.
• Verifies post-execution output existence, computes and stores checksums, and examines data for numerical anomalies.
• Emits structured JSONL audit records capturing all relevant state and QC outcomes.

This layered approach to checkpointing permits full pipeline reruns, stepwise re-verification, and fine-grained debugging. Downstream researchers or auditors can examine or rerun any stage—using the manifest—without external intervention or ad hoc reconstruction. Every decision, dependency, and artifact is persistently logged and referenceable (Wang et al., 27 Apr 2026).

5. NeuroBench: Executability and Reproducibility Benchmark

NeuroClaw is complemented by NeuroBench, a system-level benchmark tailored for neuroimaging pipeline executability, artifact validity, and reproducibility readiness. NeuroBench comprises 100 tasks across four modules:

• Basic utilities (BIDS validation, format conversion)
• Tool pipelines (ANTs alignment, FSL/FreeSurfer preprocessing)
• Multimodal workflows (structural-to-functional registration, tractography)
• Closed-loop research execution (iteration readiness, QC reporting)

Each task specifies precise assumptions, outputs, file conventions, and intermediate checkpoints. A fixed GPT-5.4 reviewer evaluates each task along three axes: Planning completeness (P), Reasonableness of tool/skill usage (R), and Command or code correctness (C). The aggregate score (on a 0–100 scale) is computed as

$$S_{10} = 0.30\,P + 0.40\,R + 0.30\,C, \quad S_{100} = 10 \times S_{10}$$

Performance is compared as mean scores with and without NeuroClaw skills, with normalized gains

$$g = \begin{cases} \Delta/(100 - S_{\text{no}}) & \Delta \geq 0 \\ \Delta/S_{\text{no}} & \Delta < 0 \end{cases}$$

where $\Delta = S_{\text{with}} - S_{\text{no}}$, and results are clipped to $[-1,1]$ [(Wang et al., 27 Apr 2026), Eq. 1].

6. Empirical Impact and Evidence

NeuroClaw’s integration with ten leading multimodal LLMs, including Claude-Opus-4.6, Claude-Sonnet-4.6, Gemini-3 Flash, GPT-5.4, and Qwen3-plus, results in consistent score improvements across all tested configurations. The average absolute gain is 4.74 points on a 0–100 scale; for instance, Claude-Sonnet-4.6 improved from 65.37% to 70.39% ($\Delta=5.02$, $g=0.145$), and MiniMax-M2.7 achieved the largest gain of 12.97 points ($g=0.20$). All ten models exhibited positive normalized gains, with notable enhancements in planning completeness and command correctness. Token consumption rises slightly, as models must parse more skill documentation and verify states, but the improvement in reliability and scientific validity is marked (Wang et al., 27 Apr 2026).

Qualitatively, NeuroClaw modularizes monolithic pipelines into a loop of planning, execution, verification, and refinement. By basing every decision on BIDS metadata and explicitly managing all dependencies, it transforms raw DICOM repositories into analysis-ready, reproducible artifacts with a single command invocation. In tandem with NeuroBench, this framework offers both a tool and a benchmark for robust, agentic neuroimaging research.