The discipline is what makes this usable for real work: model files, runs, QA checks and provenance stay on disk as artefacts you can re-run and audit, so an agent-driven run is reproducible rather than a chat transcript. The published case where the model named its own errors as disqualifying, rather than reporting numbers from a dry week, is the behaviour to look for.
A verification-first framework for reproducible stormwater modelling. You describe a modelling goal in English; model construction, SWMM execution, QA checks and reporting run as deterministic, inspectable steps that leave artefacts behind. The MCP servers and skills work with the project's own runtime or with Codex, Claude Code, OpenClaw and Hermes.
- Modelling operations exposed both as tools and as CLI verbs on the same engine: `run_swmm_inp` runs a model, `audit_run` produces the provenance record, note and diagnostics, `review_run` checks a run against the rulebook
- Calibration and scenarios — `swmm_calibrate_search` and `swmm_calibrate_sceua` against an observed CSV, `run_climate_scenarios` to compare the model under precipitation-scaled scenarios
- Inputs and outputs — `generate_design_storm` for a design rainfall series, `plot_run` for hydrographs, `map_run` for the network map, `generate_report` for a client-ready document
- Every run leaves reusable artefacts: the SWMM input file, `.rpt` and `.out` outputs, manifests and command traces, QA summaries and parsed peak-flow metrics, plus `experiment_provenance.json`, `comparison.json` and `experiment_note.md`
- Real networks upstream — SWMMCanada covers 35 Canadian cities, SWMManywhere synthesises elsewhere
- Skills install into any skills-aware runtime with `npx skills add Zhonghao1995/agentic-swmm-workflow`
One-line installers for macOS and Linux (`curl -fsSL https://aiswmm.com/install.sh | bash`) or Windows PowerShell, then `aiswmm` to launch the runtime; a pinned Docker image runs it with no local install. `pip install aiswmm==0.9.3` is the third path. After install, `aiswmm setup` lists ten LLM provider routes and detects what is already running — three need no API key, including a local gateway. Never paste API keys into the conversation; store them as documented. Pre-1.0, stable at v0.9.3. MIT.
One command plus a key — pip install aiswmm, then supply credentials