MCP server that lets an LLM agent debug regional ocean (MOM6 + CESM) model runs by inspecting log/parameter/diagnostic files and querying an expert-interview knowledge base.
Log triage: read_run_log (recent CESM/MOM6 log lines per component), find_errors (scan all logs for FATAL/ERROR/NaN/CFL/negative thickness/OOM), classify_error (match error text to known MOM6 failure modes with cause and fix suggestions).
Parameter inspection: read_mom6_params (MOM_input + MOM_override as KEY = VALUE, with overrides flagged), diff_params (side-by-side comparison of two cases), get_parameter_advice (expert guidance on a specific parameter).
Stability checks: check_cfl (sanity-check DT/DT_THERM vs. apparent resolution; caveat: no real grid dx/dy), suggest_timestep (recommended DT and DT_THERM for a given resolution).
Diagnostics: read_diag_table (parse enabled MOM6 diagnostics by module), suggest_diagnostics (recommend fields for symptoms like temperature drift, OBC instability, tidal mixing, BGC).
Knowledge base search: query_domain_knowledge (top-k expert Q&A pairs on OBC setup, tides, bathymetry, stability, BGC, grid design), backed by a keyword search over private interview transcripts.
Resources: ocean://case/{case_dir*}/params, ocean://case/{case_dir*}/errors, ocean://knowledge/index for direct URI access to the same data.
Note: the README also lists check_obc_forcing_status (CrocoDash OBC/IC/WW3/CICE forcing check) and read_cesm_timing (CESM timing/throughput summary), but neither appears in the supplied server schema.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@regional-ocean-debugger-mcp Find errors in my croc case logs"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
MCP server for debugging regional ocean (MOM6 + CESM) model runs.
Combines practical log analysis tools with a searchable knowledge base of expert interview transcripts, giving an LLM agent both the ability to inspect model output files and access to domain-expert guidance.
Tool | Purpose |
| Read recent CESM/MOM6 run log lines |
| Scan all logs for FATAL/NaN/CFL/etc. |
| Match error against known MOM6 failure modes |
| Check CrocoDash forcing outputs: OBC raw data + |
| Read MOM_input + MOM_override |
| Compare MOM_input between two cases |
| Evaluate CFL stability from MOM_input timestep settings |
| Recommend DT and DT_THERM for a given resolution |
| Parse enabled MOM6 diagnostics |
| Suggest useful diagnostics for a given problem |
| Search expert interview knowledge base |
| Look up guidance for a specific MOM6 parameter |
| Summarise CIME's |
Related MCP server: log-mcp
URI | Content |
| MOM_input + MOM_override as KEY = VALUE (same as |
| Error lines from the latest run's logs (same as |
| Summary of knowledge base contents |
{case_dir*} takes the case path as-is, slashes included, e.g.
ocean://case//glade/u/home/<user>/croc_cases/my_case/errors.
Domain knowledge comes from expert interviews on regional MOM6 modelling (private repo:
AidanJanney/RegionalOceanModelingInterviews). Topics include:
OBC setup, tidal forcing, bathymetry, stability, BGC, grid design, and regional ocean dynamics.
The transcripts are a git submodule in data/transcripts/. After cloning this repo:
git submodule update --init data/transcriptspip install -e .
# with the transcripts submodule:
git submodule update --init data/transcriptsregional-ocean-debugger-mcp
# or
python server.pyAdd to .mcp.json:
{
"mcpServers": {
"ocean-debugger": {
"command": "python",
"args": ["/path/to/MCP_regional_ocean_debugger/server.py"]
}
}
}Stateless: reads from the filesystem per call
No ML dependencies: knowledge search is keyword-based (fast, transparent, works without GPU)
Composable: use find_errors → classify_error → query_domain_knowledge in sequence for triage
Private data: transcripts submodule is private; do not make public without authorisation
Cases: ~/croc_cases/
Run dir: the case's RUNDIR (default $CIME_OUTPUT_ROOT/$CASE/run, read from env_run.xml)
MOM_input: RUNDIR/MOM_input or <case_dir>/Buildconf/momconf/MOM_input
diag_table: RUNDIR/diag_table or <case_dir>/Buildconf/momconf/diag_table
Evaluate CFL stability condition for a MOM6 case.
Reads DT, DT_THERM, and resolution-related parameters from MOM_input. Compares against recommended safe values for the apparent resolution. Returns a stability assessment and flags potential violations.
Note: A full CFL check requires the actual grid file (dx, dy values). This tool uses MOM_input parameters as a proxy for a quick sanity check.
Inputs: case_dir (required). Returns: result
Classify a MOM6/CESM error message and return probable cause and fix suggestions.
Pass a snippet of error text (e.g. from find_errors output) and get back a structured diagnosis. Matches against a library of known MOM6 failure modes.
Inputs: error_text (required). Returns: result
Compare MOM_input parameters between two CESM/MOM6 cases.
Returns parameters that differ, with values from both cases side by side. Useful for diagnosing why two configurations behave differently.
Inputs: case_dir_a (required), case_dir_b (required). Returns: result
Scan all run logs in a CESM case for error indicators.
Searches for FATAL, ERROR, NaN, CFL violations, negative thickness, OOM, and similar. Returns a structured list of matches with filename, line number, and error type. Call classify_error() on the error text for probable cause and fix suggestions.
Inputs: case_dir (required). Returns: result
Look up expert guidance for a specific MOM6 parameter name.
Searches the interview knowledge base for mentions of the parameter, along with context about what it does, typical values, and pitfalls.
param_name: MOM6 parameter name (e.g. 'DT', 'MINIMUM_DEPTH', 'OBC_SEGMENTS', 'SPONGE', 'TIDES', 'KD_MAX', 'KHTH').
Inputs: param_name (required). Returns: result
Search expert interview transcripts for scientific guidance on a question.
The knowledge base contains interviews with regional ocean modelling experts covering topics like: MOM6 configuration, OBC setup, tidal forcing, bathymetry, stability, BGC, grid design, known failure modes, and regional ocean dynamics.
Returns the top matching Q&A pairs from the expert interviews. top_k: number of results to return (default 5, max 10).
Inputs: question (required), top_k. Returns: result
Read and parse the MOM6 diag_table from a CESM case.
Returns enabled diagnostics grouped by module/component. diag_table is typically in case_dir/run/ or case_dir/Buildconf/momconf/.
Inputs: case_dir (required). Returns: result
Read MOM_input and MOM_override parameter files for a CESM/MOM6 case.
Returns all parameters as KEY = VALUE pairs. Pass filter to show only parameters containing that substring (e.g. 'DT', 'OBC', 'SPONGE'). MOM_override values take precedence and are marked with [override].
Inputs: case_dir (required), filter. Returns: result
Return recent lines from a CESM or MOM6 run log.
component: 'cesm', 'ocn' (MOM6), 'atm', 'ice', 'rof', or 'all'. Searches case_dir and case_dir/run/ for log files. Returns the most recent match.
Inputs: case_dir (required), component, lines. Returns: result
Suggest useful MOM6 diagnostics to add to diag_table based on a description of the problem.
symptoms: describe what you're investigating (e.g. 'temperature drift', 'OBC instability', 'CFL violations', 'tidal mixing', 'BGC', 'mixed layer too shallow').
Returns a list of recommended diagnostic fields grouped by relevance.
Inputs: symptoms (required). Returns: result
Suggest safe MOM6 timestep values (DT and DT_THERM) for a given grid resolution.
resolution_deg: approximate horizontal resolution in degrees (e.g. 0.1 for 1/10°). Returns recommended values with caveats about bathymetry and tidal forcing.
Inputs: resolution_deg (required). Returns: result
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 11 tools
Each tool has a clearly distinct purpose, ranging from CFL checks to error classification, parameter comparison, and diagnostics suggestions. No two tools overlap in functionality.
All tool names follow a consistent verb_noun snake_case pattern (e.g., check_cfl, find_errors, suggest_diagnostics) with no deviations.
Eleven tools is an appropriate number for a specialized debugger, covering core tasks without being excessive or insufficient.
The tool set covers the main debugging workflow (error detection, classification, parameter analysis, diagnostics advice) but lacks a tool for direct parameter modification, which is a minor gap.
MCP server for the Fail Modes taxonomy — a knowledge base of AI system failure modes
MCP server for querying Forkast documentation
MCP observability: analyse MCP server usage, investigate sessions, and instrument MCP servers.
