Coral

🪸 CORAL — Coastal Ocean Research AI Layer

A self-hosted AI agent for NOAA HPC that connects local LLMs to ocean data,scientific documentation, and HPC workflows entirely within NOAA's network.

What It Does

• Ocean data & analysis — Real-time water levels, hurricane tracks, storm surge forecasts, satellite data via ocean-mcp, plus Python code execution for plotting and analysis
• Code & documentation — RAG search over SCHISM/ADCIRC source code, NOAA tech memos, namelists, and NOS workflow configs
• HPC workflows — Slurm and PBS job diagnostics, ecFlow suite monitoring, UFS-Coastal experiment management, disk quotas, FairShare, threshold alerting
• Persistent & portable — Memory across sessions, CLI with slash commands, web UI, deployed on NOAA Ursa and TACC Vista

All self-hosted on Ollama with open-weight LLMs. No external APIs, no cloud dependencies.

Quick Start

git clone https://github.com/mansurjisan/coral.git
cd coral
pip install -e .
ollama pull qwen3:32b
coral chat --model qwen3:32b --mode multi

For HPC deployment, see [NOAA Ursa setup](docs/setupursa.md) or [TACC Vista setup](docs/setupvista.md).

Supported Models

CORAL works with any model available in Ollama. Pull the model and pass it with --model:

ollama pull gemma4
coral chat --model gemma4 --mode multi

You can also switch models mid-session using the @ prefix:

You: @gemma4 What is the current water level at The Battery?

| Model | Size | Notes | |-------|------|-------| | qwen3:32b | 20 GB | Default, strong tool calling | | gemma4 | 10 GB | Google's latest, good reasoning | | llama3.3:70b | 40 GB | Largest open model, needs >48 GB VRAM | | qwen3:8b | 5 GB | Lightweight, fast, good for testing |

MCP Servers

CORAL connects to 22 MCP servers providing 150+ tools across three categories:

• Ocean data (12 servers) — CO-OPS, NHC, STOFS, ERDDAP, OFS, GOES, USGS, NDBC, WW3, ADCIRC, SCHISM, Hurricane Recon via ocean-mcp
• HPC & workflow (6 servers) — Slurm, PBS (WCOSS2), ecFlow, UFS experiment runner, HPC system admin, NOS workflow configs
• Local tools (4 servers) — NetCDF queries, RAG documentation search, Python execution (sandboxed), threshold alerting

CLI Features

╭──────────────────────── CORAL v0.1.0 ────────────────────────╮
│  🪸 Model  qwen3:32b  │  Tools  150  │  Mode  multi-agent   │
│  🖥️  User   mansurjisan  │  Host   login2.vista.tacc.utexas.edu │
│  🧠 3 memories loaded                                        │
╰──── Coastal Ocean Research AI Layer · /help for commands ────╯

| Command | Description | |---------|-------------| | /help | Show all commands | | /tools | Tool count per section (DATA/CODE/WORKFLOW) | | /status | Quick dashboard: Ollama health, running jobs, disk usage | | /memory | Show saved memories | | /remember | Save a preference (e.g. /remember account = coastal) | | /forget | Remove a memory | | /save | Export conversation to markdown | | /report | Auto-generate HPC status report | | /techmemo | Auto-generate NOAA tech memo draft | | /alert | Set threshold alert (e.g. /alert 8518750 > 1.5) | | /watch | Monitor a job (e.g. /watch 9848988) | | /branch | Save conversation, start fresh | | /audit | Show tool call history and stats | | /clear | Clear conversation history | | @model | Override model for one query (e.g. @qwen3:8b What is SCHISM?) |

Audit & Policy

• Audit logging — Every query and tool call recorded with timestamps, routing decisions, tool timing, confidence scores, and sandbox usage
• Policy manifest (src/coral/policy_manifest.json) — Defines which MCP servers each agent section can access, trust classes, and per-environment sandbox requirements
• Sandbox enforcement — Set CORAL_REQUIRE_SANDBOX=1 to block host-side Python execution
• Tool caching — 5-minute TTL cache for read-only tools (HPC status, configs)
• Retry with backoff — Automatic retry on transient Ollama connection failures

Related

• ocean-mcp — MCP servers for NOAA ocean data (18 servers)
• nos-workflow — NOS Unified Operational Forecast System workflow
• Ollama — Local LLM inference
• Model Context Protocol — Tool integration standard

Citation

If you use CORAL in your research or operations, please cite:

@software{jisan2025coral,
  author = {Jisan, Mansur},
  title = {CORAL: Coastal Ocean Research AI Layer},
  year = {2025},
  url = {https://github.com/mansurjisan/coral},
  note = {A self-hosted AI agent connecting local LLMs to NOAA ocean data via MCP}
}

Author

Mansur Jisan — NOAA National Ocean Service

License

Apache 2.0

Source & license

This open-source MCP server is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: mansurjisan
• Source: mansurjisan/coral
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.