Skip to content

README.md

Latest commit

 

History

History
82 lines (63 loc) · 3.38 KB

README.md

File metadata and controls

82 lines (63 loc) · 3.38 KB
Raw

MCP_hub

Single entry point for the CROCODILE-CESM MCP servers. Wires up all three so an MCP client (Claude Desktop, Claude Code, Cursor, etc.) can load them together, and vendors the source of each as a submodule for offline/HPC use.

The servers

Server Repo Role
crocodash MCP_CrocoDash Configure and deploy regional MOM6 ocean models within CESM (create/process/bundle a case)
cesm-runner MCP_cesm_runner Drive a CESM case through setup, build, submit, and monitoring (bare-metal or container)
regional-ocean-debugger MCP_regional_ocean_debugger Post-run diagnostics and scientific guidance for MOM6/CESM runs

They're designed to be used together across a run's lifecycle: crocodash creates and configures a case, cesm-runner builds/submits/monitors it, and regional-ocean-debugger helps diagnose the result. See each server's own server.py docstring for the exact deployment paths (HPC batch queue, HPC + container, laptop + container).

Quick start (recommended): remote, via uvx

No local checkout needed. Requires uv and outbound internet access at launch time. Point your MCP client's config at this repo's mcp.json, or copy its mcpServers block into your client's own config file.

Offline / HPC setup

Compute nodes often lack outbound internet access, so uvx --from git+... won't work there. Instead, vendor and install all three servers locally:

git clone --recurse-submodules https://github.com/CROCODILE-CESM/MCP_hub.git
cd MCP_hub
./setup.sh

This creates a single venv, installs all three servers into it (editable), and generates mcp.local.json (git-ignored, since it contains absolute paths) pointing at the installed console scripts. Point your MCP client at mcp.local.json instead of mcp.json.

crocodash-mcp needs ESMF/xesmf, which are conda-only compiled libraries a plain venv can't provide. If you already have a conda env with those installed (e.g. whatever env you use for the CrocoDash CLI itself), point PYTHON_BIN at a clone of it rather than the env itself:

conda create --name mcp-hub --clone CrocoDash
PYTHON_BIN=/path/to/envs/mcp-hub/bin/python ./setup.sh

Prefer a clone over installing directly into your working env: the servers' dependencies (auth/keyring/crypto libraries pulled in by fastmcp) are extra weight you likely don't want mixed into an env you use for other work, and a clone means an update or cleanup of one env can't affect the other.

Updating the vendored servers

The submodules are pinned to specific commits. To bump all three to the latest main:

./update-submodules.sh

Review the diff, then commit the updated submodule pointers.

Layout

mcp.json                   # default config: remote servers via uvx
mcp.local.json.template    # template for offline config, filled in by setup.sh
setup.sh                   # one-time offline/HPC setup (venv + editable installs)
update-submodules.sh       # bump vendored servers to latest main
servers/
  crocodash/                  # submodule -> MCP_CrocoDash
  cesm-runner/                # submodule -> MCP_cesm_runner
  regional-ocean-debugger/    # submodule -> MCP_regional_ocean_debugger