prahaladuk2208-arch/geeflow

geeflow lets Claude control Google Earth Engine, a service that stores and processes huge amounts of satellite imagery on Google's own servers. It works through the Model Context Protocol, a standard that lets AI assistants call external tools. With geeflow connected, you can ask Claude to build a cloud free satellite picture of any location on Earth, calculate vegetation health measurements, pull statistics for a region, or classify what kind of land cover is present, and Claude actually carries this out on Google's infrastructure rather than just describing how to do it. The tools it exposes include searching Google's catalog of over 880 satellite datasets, building cloud free composite images for any place and year back to 1972, computing spectral indices used to measure plant health or water presence, generating preview images, pulling per region statistics, creating labeled training data from existing land cover maps, exporting results to Google Drive or an Earth Engine asset, and tracking those exports. A general purpose code tool covers anything the typed tools do not, such as climate or terrain analysis. To use it you need Python 3.10 or newer, a free Google Earth Engine account, and a Google Cloud project with the Earth Engine API turned on. All the actual computation happens on Google's servers, so there are no large downloads and no need for a powerful graphics card on your own machine. Setup involves installing the package, authenticating with Earth Engine through a browser window, and adding the server to Claude Code or another compatible AI client with a short command. Underneath the AI connector is a regular Python library and command line tool, so anything Claude does through geeflow, a person can also do with plain Python code or direct commands, without an AI involved. There is also an optional add on package for land cover classification, where one configuration file describes a whole project, including the area, years, and land classes, and the tool compares several machine learning models before producing a map. The project emphasizes reproducibility and honest results: classification accuracy is measured using a method that keeps pixels from the same training area together during testing, avoiding an inflated accuracy score, and every workflow can be rerun later from its saved configuration file. It includes 69 automated tests that run without an Earth Engine account. It is released under the MIT license.