Skip to content

MESA RStudio Geospatial

Repo: idss-mesa/rstudio · Image: harbor.cyverse.org/vice/mesa-rstudio:latest (GPU: :gpu) · In the portal: Applications → MESA Apps → MESA RStudio Geospatial

RStudio Server on the Rocker geospatial stack, with the MESA AI coding agents and CyVerse Data Store tools added1. There is no RStudio login: CyVerse signs you in when the app opens.

What's inside

Category Tools
IDE RStudio Server
Science R with the tidyverse, sf, terra, and stars; GDAL, PROJ, and GEOS
AI agents and MCP Claude Code, Codex, OpenCode, Antigravity, Claude Code Router; the irods, mesa, formation, and filesystem MCP servers — see AI agents in the MESA apps
CyVerse data GoCommands, iRODS configuration, S3/OSN mounts, AWS CLI
Developer tools GitHub CLI, Git Credential Manager, Go 1.25, Node.js 22

The agent CLIs are on the PATH both in RStudio's Terminal tab and in R, for example system("claude --version").

Start it

  1. In the MESA Portal, open Applications → MESA Apps.
  2. On MESA RStudio Geospatial, click Instant Launch or Launch with Options. See Starting applications.
  3. RStudio opens in a new tab with ~/data-store (your Data Store) as the working folder.

You can also start it from the Discovery Environment by searching for MESA RStudio Geospatial.

First steps

  1. Open the Terminal tab in RStudio.
  2. Run cyverse-login to give the tools and agents your CyVerse access.
  3. Optionally run aiverde-setup to connect AI Verde models.
  4. Start an agent in the terminal: claude, codex, opencode, or agy.

See AI agents in the MESA apps. Save scripts, projects, and outputs under ~/data-store; the rest of the container is deleted when the analysis ends.

GPU build

harbor.cyverse.org/vice/mesa-rstudio:gpu adds GPU computing to the same workbench. It adds:

Adds Details
Deep learning in R torch with its CUDA runtime, luz, torchvision, tabnet, brulee, tidymodels
GPU xgboost xgboost built for the GPU (device = "cuda"); lightgbm on the CPU
Local LLMs An Ollama server on the GPU, with the ollamar, ellmer, and mall R packages — see Local models on a GPU
Python from R reticulate, keras3, and tensorflow; the first library(keras3) downloads TensorFlow and its CUDA libraries (about 13 GB, about a minute; kept until the analysis ends)
GPU tools mesa-gpu-check, nvtop, nvitop

Try it in R:

library(torch); cuda_is_available()           # TRUE on a GPU node
ollamar::generate("qwen3.5:9b", "Summarise this abstract: ...", output = "text")   # after ollama-setup

Two cautions:

  • install.packages("torch") or install.packages("xgboost") replaces the GPU builds with CPU-only ones. Leave those two packages as they are.
  • Load keras3 or tensorflow before library(torch) in a session; R torch and Python torch cannot be loaded in the same R session.

Run it on your own computer

docker run --rm -p 8787:80 -e IPLANT_USER=$USER -e REDIRECT_URL=http://localhost:8787 \
  harbor.cyverse.org/vice/mesa-rstudio:latest

Open http://localhost:8787. If a new browser lands on a "not found" page, open http://localhost:8787/auth-sign-in once and then http://localhost:8787/ again. Outside CyVerse there is no password, so publish the port only on your own machine. The image is built for linux/amd64.

  1. MESA RStudio Geospatial README, https://github.com/idss-mesa/rstudio. ↩

Machine-readable versions of this page: Markdown twin · raw source on GitHub · llms.txt · llms-full.txt (whole site). See For AI agents.