Follow-ups stay with their model
Questions return to the run they refer to. A failed fetch remains visible as a failed fetch, and provider errors leave the interactive session available to continue.
Latest v0.9.4 · lessons from five cities ↗
Open source · verification first
Stormwater modelling with a chain of evidence.
Describe the job in a sentence. Fetch a municipal storm network from SWMMCanada, run EPA SWMM, inspect uncertainty, and export an audited report with its inputs, checks, and limits in view.
- v0.9.4 stable
- MIT licensed
- EPA SWMM 5.2.4
- PyPI + Docker
58typed tools
19reusable skills
4,067tests reported in v0.9.4
10LLM provider routes
Current release · September 2026
v0.9.4 keeps the answer tied to the run.
Two days of natural-language sessions across five Canadian cities shaped this release: stronger run context, new analysis tools, and clearer evidence in every deliverable.
Read the changelog →Explore uncertainty before observations arrive
Propagate stated parameter ranges or rank their effects one at a time. Each analysis retains its own files and reports flow in the model’s units, with its assumptions stated.
Past failures inform the next attempt
Failure memory loads at session start, relevant parameter records reach the planner, and new lessons stay anchored to the place being modelled.
Reports carry the limits of the result
Word reports state evidence boundaries and any design-review verdict. Re-audits retain the previous record, and expert decisions record an explicit yes or no.
The method
Ask naturally. Execute deterministically. Keep the record.
This is not a chat wrapper around SWMM. The agent coordinates; the solver, checks, and artifacts remain inspectable.
Source or build the model
Fetch a real municipal storm network through SWMMCanada, the upstream model-building project in the same ecosystem—or bring an existing .inp or prepared GIS layers.
Run, calibrate, and verify
EPA SWMM runs stay CLI-runnable. QA gates, SCE-UA calibration, precipitation-scaled climate scenarios, and direct solver comparisons keep automation grounded.
Audit, remember, refine
Provenance and experiment notes feed curated and raw memory. Skill changes remain proposals until a human reviews and benchmarks them.
New case · downtown Kelowna, BC
One sentence. One approval. A complete record.
The fifth city in the release campaign follows the same fetch, run, audit, and report chain. The model, figures, report, and session transcript are published together, so the SWMM simulation can be repeated offline.
Explore the Kelowna evidence →A published storm network
411 derived subcatchments, 889 storm nodes, 832 conduits, and 76 outfalls. The bundle also contains sanitary pipes; this simulation uses the storm system alone.
Numbers with a source
Principal-outfall peak 0.130 m³/s; system peak outflow 0.706 m³/s. Runoff continuity error −0.078%, routing −1.022%. The audit passed all three checks.
Geometry warnings remain visible
Ten conduits run uphill and five subcatchments route more than 50 m to their outlet. The model is an uncalibrated first pass with no observed flow supporting its predictions.
The deliverable and its inputs travel together
The published case includes the original model.inp, a sample Word report, upstream QA, report excerpts, and the full two-turn transcript.
The upstream
SWMMCanada builds the model. Agentic SWMM runs, audits, and defends it.
SWMMCanada is the data-and-model-building layer of the same ecosystem: draw an area, and it assembles a ready-to-run model.inp from Canadian open data. Agentic SWMM consumes it through one typed tool and carries the retrieved archive and service identifiers into the run record.
Evidence, not theatre
Every visual has a stated boundary.
Benchmarks, research previews, and audit examples are presented with what they prove—and what they do not.
Kelowna / upstream bundle
Real pipes, derived subcatchments, recorded assumptions
Map includes both networks; simulation uses storm pipes
Kelowna / executed run
Rainfall and the principal outfall’s response
Uncalibrated model, 1–4 November 2023
Audit example
A mismatch the workflow refuses to hide
Metric provenance check
Project overview
See the workflow, then inspect the implementation.
The short introduction explains the modelling loop. The repository contains the runtime, 19 skills, 11 MCP servers, 58 typed tools, tests, benchmarks, and audit contracts.
Published research · AI for Engineering · 2026
Auditable and reproducible stormwater modelling with Agent Skills and MCP.
Zhang, Z. & Valeo, C. · Volume 1, Issue 1, Article 5
Read the paper ↗Go deeper
Follow the question you have.
01
Why this exists
The modelling problem and the human-control boundary.
Vision → 02How it works
Runtime, skills, evidence contracts, memory, calibration, and climate scenarios.
Features → 03What is supported
Benchmarks, research previews, and explicit evidence limits.
Validation → 04How to run it
One-line installer, PyPI, Docker, provider routes, and portable skills.
Install →