Agentic MIKE+

A headless, natural-language-driven, automated modelling workflow for MIKE+. State a goal in plain language; an agent inspects the model, changes parameters, runs simulations, reads results, and plots — end to end, without ever opening the MIKE+ GUI.

Experimental / pre-release. One MCP server + skills wrapping DHI's Python stack (mikeplus / mikeio / mikeio1d). Verified end to end on the MIKE+ 2026 Sirius_RTC example. Sibling of agentic-swmm-workflow.

Quick start

Windows + Claude Code — one command. Sets up a Python 3.11 venv, installs the pinned deps, and registers the MCP server with Claude Code:

powershell -ExecutionPolicy Bypass -File scripts\install.ps1

By hand (or for another MCP client):

py -3.11 -m venv .venv
.\.venv\Scripts\python.exe -m pip install -r requirements.lock
.\.venv\Scripts\python.exe -m pip install -e .
claude mcp add mike-plus -- "<repo>\.venv\Scripts\python.exe" -m mikeplus_mcp.server   # Claude Code

Needs Python 3.11 (x64); run/edit also needs a licensed MIKE+ 2026. For other MCP clients, copy config/mcp.sample.json and fix the paths. Versions are pinned in requirements.lock.

Related MCP server: SWMM-MCP

Why it matters

• Natural-language-driven. Describe the task in plain words; the agent plans and runs it — no scripting, no GUI clicking. (A sub-agent did this autonomously.)

• Fully headless. Runs with no GUI — on a workstation, a server, in CI, or under an agent. Built for batch and scenario automation.

• MCP-native and portable. One server speaks the Model Context Protocol; connect Claude Desktop or Codex with a single config line, and install with pip.

• Low barrier to share. Reading results and plotting need no MIKE+ license — only running or editing does. Teammates analyse model output with nothing but pip install.

• Reproducible. Deterministic tools, structured-JSON output, a pinned lockfile, verified on a real model — not a chat-to-model black box.

• Engine-agnostic and extensible. Results use a common schema (ready to sit beside SWMM and LSTM); add a tool or skill by dropping in a file.

How it works

Skills (markdown playbooks) tell the agent when and how; the agent calls MCP tools; each tool runs in an isolated worker subprocess that imports only mikeplus or mikeio* — the two cannot share a process. The server itself imports neither.

agent  ->  reads skills/*.md  ->  calls MCP tools  ->  workers (mikeplus / mikeio1d)

Tools (one server, mike-plus)

Five skills (mike-model, mike-params, mike-runner, mike-results, mike-plot) orchestrate them.

Demo — Sirius_RTC (MIKE 1D, 568 nodes, 576 links)

Full evidence — commands, outputs, and the honest license boundary — is in docs/verification.md.

License

MIT © 2026 Zhonghao Zhang, University of Victoria. Built on DHI's mikeplus / mikeio / mikeio1d and the Model Context Protocol. MIKE+ is a product of DHI; running models requires a valid DHI license.