HF radar operators run the same routine reviews over and over: is each site still delivering radials, has coverage changed compared to last week or last month, does the radial data look wrong, and when a site goes quiet, why. Most of this is done by eye, and the results usually end up as a note in a log rather than something reproducible.

Much of it can be computed directly from the radial files and ingest logs an RA already holds. This project proposes two connected pieces:

1. An MCP server for HFR network state. MCP (Model Context Protocol) is the emerging standard for exposing tools to AI agents. The idea is to expose the routine checks as a small set of read-only tools, for example:

  • site list and last radial timestamp per site (is it online, is data flowing, how delayed)
  • radial coverage/distribution comparison between two periods — week vs. week, month vs. month, or any custom window
  • radial quality signatures: new coverage gaps, reduced vector counts, out-of-bound vectors at unexpected bearings, opposing vectors in adjacent range bins, shortened range
  • QARTOD flag summaries per site over a time window
  • recent sync/ingest history

Once these exist, any agent — Claude, ChatGPT, or a local model — can query the network, and the same tools can also be called directly from a script or notebook without an LLM involved.

2. A lightweight operations agent built on those tools. Two uses:

  • Dropout diagnosis. When a site stops reporting, the agent runs the checks above and writes a plain-language summary, e.g. “SGRV: no new radials since 09-20 14:00. Other sites on the same server are current, so this looks site-side rather than a sync failure. Vector counts had been declining for three days before the dropout.”
  • Periodic review narrative. Run the coverage comparison on a schedule and have the agent draft the write-up plus a dated log entry, so a review that’s normally a visual check becomes a reproducible record with numbers behind it.

Read-only tools only. The agent reports and drafts; it does not act on the network.

Scope note: this covers what can be derived from radial data and ingest logs. Site-side diagnostics (hardware, spectra, antenna patterns) need access to the radar site computers and are deliberately out of scope here, though the tool interface should leave room for someone to add them later.

GCOOS HFR would be the reference implementation (CODAR SeaSonde .ruv radials, ~10 sites across two partner servers), but the tools should be generic enough for another RA to point at their own radial archive or at the national network.

Expected Outcomes

  • A working, read-only MCP server for HFR network status (repo + install instructions)
  • A documented set of tool definitions other RAs can adapt
  • Example agent sessions diagnosing real past station dropouts, as notebooks or transcripts
  • A sample auto-generated periodic review, showing what a reproducible version of a routine coverage comparison looks like
  • A small evaluation set of known past incidents, to check whether the agent’s diagnosis is actually right
  • Short write-up on what worked, what didn’t, and where agents are the wrong tool

Skills required

  • Python
  • Anyone who operates an HFR network or does QARTOD QC, or interest in them
  • Interest in MCP / AI agents; no prior experience needed

Expertise

Intermediate

Topic Lead(s)

Xiao Qi (GCOOS) — co-leads welcome, especially from other RAs running HFR

Relevant links

Functioning Prototype

TBD

Workflow

Tasks:

Updates