Agentic-MIKE-Plus

Agentic-MIKE-Plus is a headless, natural-language-driven automated modeling workflow for MIKE+ urban water modeling. It consists of Skills (Markdown operation manuals) and an MCP server, and can be used with AI coding agents such as Claude Code, Codex, Hermes, or OpenClaw: describe the goal in natural language, and the agent can inspect, edit, run, read, and plot the MIKE+ model, completing modeling and analysis end to end without ever opening the GUI.

Purpose and Positioning This project is aimed at MIKE+ modeling tasks that require batch and scenario automation, and can run on workstations, servers, CI environments, or under an agent. It is marked as experimental/pre-release, has been validated end to end on the Sirius_RTC example of MIKE+ 2026, and is a sister project of agentic-swmm-workflow.

Main Capabilities

• Natural language driven: describe tasks in plain language, and the agent plans and executes on its own, without writing scripts or clicking through the GUI.
• Fully headless: does not depend on the GUI, making it suitable for batch processing and scenario automation.
• MCP-native and portable: one server implements the Model Context Protocol, can work with Claude Code, Codex, Hermes, or OpenClaw through a one-line configuration, and is installed via pip.
• Low barrier to sharing: reading results and plotting do not require a MIKE+ license; only running or editing does. Team members can analyze model output with just pip install .
• Reproducible and tested: deterministic tools, structured JSON output, a pinned-version lockfile, and a license-free unit test suite in CI, validated on real models.
• Engine-agnostic and extensible: results use a generic schema (can be listed alongside SWMM and LSTM); tools or skills can be added by dropping in a single file.

Architecture and Tools Skills (Markdown operation manuals) tell the agent when and how to operate; the agent calls MCP tools; each tool runs in an isolated worker subprocess that only imports mikeplus or mikeio* (the two cannot share the same process). The server itself does not import either of these.

The server is named mike-plus and is orchestrated by ten skills: mike-end-to-end (SOP: modes, run directory, stop rules, QA gates), mike-model, mike-params, mike-runner, mike-import, mike-results, mike-compare, mike-plot, mike-rain, mike-audit.

Installation and Usage The project requires Python 3.11 (x64) and offers two installation configurations:

• Reading and plotting: license-free, cross-platform, run pip install -e . (does not install mikeplus).
• Running and editing: requires Windows and a licensed MIKE+ 2026, run pip install -e ".[run]" (or pip install -r requirements.lock for an exactly pinned environment). Here mikeplus is an optional [run] extra, so team members who only read results or make charts do not need to install anything tied to a license.

Skills can be installed to any agent that supports skills (Claude Code, Codex, OpenCode, etc.) with a single command, without cloning the repository.

Use Cases

• MIKE+ modeling work that requires batch runs and scenario automation.
• GUI-less model checking, parameter editing, simulation runs, and result analysis in CI or server environments.
• Team collaboration where some members only need to read results and plot, without a MIKE+ license.
• Scenarios that require comparing MIKE+ results side by side with results from other engines such as SWMM and LSTM.

Limitations and Notes

• The project is in an experimental/pre-release state.
• Running or editing models requires a valid DHI license, a Windows environment, and MIKE+ 2026; reading results and plotting are license-free and cross-platform.
• The project is built on DHI's mikeplus / mikeio / mikeio1d and the Model Context Protocol; MIKE+ is a DHI product.
• Licensed under MIT (© 2026 Zhonghao Zhang, University of Victoria).
• For complete evidence (commands, outputs, and license boundaries), see docs/verification.md.
• If a runnable model is needed, SWMMCanada can build one anywhere in Canada: draw an area on the map, and a runnable stormwater model can be assembled from Canadian open data (8 cities use real municipal stormwater networks, other areas use synthetic data), providing agents with a real network they can run, edit, read, and plot; visit swmm.h2ox.me to try the hosted demo (no installation required), or visit h2ox.me to see the project.