CORAL: A Localized Multi-Agent AI System for Ocean Science Research

This article introduces the CORAL project, a self-hosted AI assistant designed specifically for NOAA ocean science research. The system addresses issues in ocean research such as massive data processing, low efficiency in tool switching, and data security. Its core capabilities include an MCP architecture connecting 12 ocean data sources, a RAG knowledge base, HPC workflow integration, and fully offline operation to meet the compliance requirements of government agencies.

Ocean science research involves massive data processing (real-time water level monitoring, hurricane path prediction, satellite remote sensing, HPC simulation). Traditional methods require switching between multiple system tools, manual data querying, and writing complex scripts, which are inefficient and error-prone. Agencies like NOAA have strict requirements for data security and network isolation, and commercial APIs pose risks of sensitive data leakage, so fully offline local AI systems have become a necessity.

The technical core of CORAL is the MCP (Model Context Protocol) architecture, which standardizes the integration of AI with external tools. It integrates 12 types of ocean data sources via 12 MCP servers: observation (CO-OPS water level stations, USGS river flow, etc.), forecasting (NHC hurricane warnings, STOFS storm surge forecasts, etc.), satellite (ERDDAP satellite data, GOES real-time imagery), and model support (ADCIRC/SCHISM configuration parsing). A total of 108 tool functions are exposed to cover end-to-end needs.

CORAL deeply integrates with HPC environments: NetCDF file processing (automatically generates xarray code to analyze model outputs), Slurm job management (query status, diagnose failure causes like OOM), ecFlow workflow monitoring, and visualization support (generates matplotlib/cartopy charts). The RAG knowledge base includes SCHISM/ADCIRC source code, NOAA technical memoranda, model configuration files, etc., supporting natural language queries for technical documents (e.g., parameter meanings).

Typical scenarios:

1. Real-time water level query: Obtain current water level and storm surge residual at The Battery station in New York;
2. Forecast comparison: Analyze the STOFS forecast peak vs. flood threshold in the next 48 hours;
3. Job failure diagnosis: Parse Slurm job exit code 137 (OOM) and provide recommendations;
4. Technical document query: Explain the control logic of the SCHISM parameter IHFSKIP.

Deployment steps: git clone code → pip install → Ollama pull model (e.g., qwen3:8b), supporting command line and Gradio Web interface. Technical highlights: fully localized deployment (compliant), open MCP ecosystem, deep integration of domain knowledge, multimodal interaction, open-source and extensible (Apache 2.0 license).

Prospects: The architecture can be extended to fields like meteorology, seismology, astronomy, bioinformatics, etc. Limitations: Limited capabilities of local open-source models, need to expand tool coverage, UI can be optimized, challenges in offline knowledge updates.

CORAL represents an important direction for AI applications in scientific research—a domain-specific professional assistant. By connecting the tool ecosystem via MCP, meeting compliance through local deployment, and integrating knowledge with RAG, it provides a practical example for ocean research. Its architecture offers a replicable blueprint for other fields, helping researchers focus on core issues.