Reproducibility-first AI infrastructure for computational hydrology and Earth system science.

Fewer than 7 % of published computational hydrology studies provide materials sufficient for independent replication. AI-Hydro exists to change that structurally rather than culturally: every tool invocation writes a run ID, a deterministic auditor gate checks every number before it enters prose, and a capsule export makes any session replay-verifiable from scratch. Reproducibility becomes a byproduct of doing the research, not a documentation step after.

We build an open, agent-native research platform and the library ecosystem behind it — so a researcher can describe their intent and receive real, audited, citable computation back.

The researcher-facing surface. An AI agent chat interface connected to your LLM of choice, with automatic detection and configuration of the aihydro-mcp server, a custom agent system prompt for hydrological reasoning, and live panels for maps, claims, evidence boards, and session replay.

VS Code Marketplace · Docs

The Python backbone. 144 validated, tiered tools exposed via the Model Context Protocol — usable from Claude, GPT, Gemini, or any MCP-compatible client, or directly as a Python library. Includes the full defensibility infrastructure: run-reference auditor gate, claims lifecycle, capsule export with replay CI, and a 60-task HydroResearch-Bench accuracy benchmark.

pip install aihydro-tools · PyPI · Tool Reference

A layered ecosystem of standalone Python packages, each independently installable, composable through a shared domain-agnostic core:

Package	What it does
aihydro-core ·	Zero-dependency substrate — HydroResult contract, bootstrap CI, ClaimStore/Auditor science protocols, uniform tool contracts
aihydro-data ·	Global data router — 54 products, 18 variables, 8 regions (USGS · GridMET · GEE · STAC · HyRiver) with carried provenance and automatic regional fallback
aihydro-watershed ·	Tiered watershed delineation (NLDI → MERIT-Hydro → pysheds), hydrological signatures, terrain analysis — any location on Earth

Catchment attribute extraction for CAMELS CONUS gauges — 70+ attributes across six categories (climate, topography, soil, geology, vegetation, land cover). Published on PyPI and archived on Zenodo.

pip install camels-attrs · PyPI · Zenodo DOI

Global GLiM lithology and GLHYMPS hydrogeology for any watershed — area-weighted from CCGM-permitted sharded GeoParquet tiles served via HuggingFace.

pip install pygeoglim · PyPI · Zenodo DOI

An agent-native pipeline that brings SWAT+ model construction under a governed, reproducible regime. The agent sequences the build steps while tested tools perform geoprocessing and configuration; runtime claim-governance gates (build · provenance · physical consistency · parameter sensitivity · calibration · routing · soil) ensure no assertion stands without machine-verifiable backing.

The platform is designed for community extension — not just of tools, but of every layer:

Each surface accepts a self-describing manifest; a shared recognition service records installs and attribution so contributed artefacts carry citable credit.

Question (natural language)
↓
AI Agent ← any MCP-compatible model (Claude · GPT · Gemini · …)
↓ JSON-RPC over stdio
aihydro-tools (144 validated, tiered tools)
↓ ↓ ↓
aihydro-data aihydro-watershed
global router delineation + signatures
USGS · GEE terrain · globally
STAC · HyRiver hydrological analysis
↓
aihydro-core (provenance · hashing · job dispatch · tool contracts)
↓
Defensibility layer
─ run ID on every result · auditor gate · claims lifecycle
─ capsule export · replay CI · defensibility report
↓
Defensible answer + evidence bundle

AI-Hydro is designed to be extended at every layer. The highest-impact contributions are new domain tools packaged as Python entry-point plugins — you don't need to fork the core.

Open domains:

→ Plugin Guide · Open an issue

_{Built for the hydrology and Earth system science community · Apache 2.0}