AI agents in the MESA apps¶

All five featured apps share the same MESA agentic stack: AI coding-agent command-line tools, the MESA MCP servers already registered with each of them, and the CyVerse Data Store tools¹. This page covers what is the same in every app; each app's page covers what is different.

What every app includes¶

The agents are command-line tools: open a terminal in the app (a JupyterLab or VS Code terminal, the RStudio Terminal tab, a terminal on the KASM desktop, or the MESA CLI itself) and type claude, codex, opencode, or agy.

Your Data Store inside the app¶

When the app starts from the portal or the Discovery Environment, your Data Store is mounted under ~/data-store, and the app opens there. Save work you want to keep under ~/data-store: everything else lives on the container's disk and is gone when the analysis ends.

At start-up each app also copies .gitconfig, .aws/, and .ssh/ from your Data Store home folder (/iplant/home/<username>/) into the app's home folder, so Git and AWS settings you keep there follow you into every session.

1. Sign in to CyVerse¶

cyverse-login          # your CyVerse username and password

cyverse-login writes the standard iRODS credentials (~/.irods/), so GoCommands, the mesa and formation MCP servers, and the agents act as you, with access to your home folder and to what is shared with you. Without it they have anonymous, read-only access to public data. Restart an agent after signing in so its MCP servers pick up the credentials.

Claude Code also registers the hosted CyVerse Data Store MCP servers: irods (the anonymous public endpoint, which works at once) and irods-auth (the authenticated endpoint). To reach your private home folder through irods-auth, sign in once per session:

claude mcp login irods-auth --no-browser   # opens a kc.cyverse.org URL; paste the redirect back

OpenCode, Codex, and Antigravity use the local servers and gocmd, which read the ~/.irods credentials that cyverse-login wrote.

2. Connect a language model¶

The images contain no API keys: each person brings their own.

CyVerse AI Verde¶

AI Verde is CyVerse's hosted LLM service². Get an API key from https://chat.cyverse.ai (Course → API Key), then in a terminal:

aiverde-setup          # paste your AI Verde key

It checks the key, lists the models your course can use, and saves the settings to ~/.config/aiverde/env (readable only by you). Then:

Your own accounts¶

Each agent can also use its vendor's own sign-in, for example claude with an Anthropic account or codex with an OpenAI account. Follow the prompts the first time you start it.

Local models on a GPU¶

The GPU builds of every app run an Ollama server inside the container, so agents can use an open model on the GPU with no API key and nothing leaving the analysis:

ollama-setup                                   # pulls qwen3.5:9b (the default) and prints these commands
ollama launch claude --model qwen3.5:9b        # Claude Code on the local model
codex --oss --local-provider ollama -m qwen3.5:9b
opencode -m ollama/qwen3.5:9b

One 16 GB NVIDIA A16 GPU fits qwen3.5:9b, gpt-oss:20b, gemma4:12b, or qwen3:4b (ollama-setup --help lists them); larger models spill onto the CPU. Models are stored in ~/.ollama/models on the container's disk and are deleted when the analysis ends. ollama stop <model> frees the GPU memory for PyTorch or other work.

mesa-gpu-check (mesa-gpu-check --ollama for a short model test) checks the GPU, the driver, CUDA, PyTorch, and Ollama.

3. Ask¶

With credentials in place, ask an agent for things such as:

• "List the folders in my CyVerse home and tell me which ones have AVU metadata."
• "Find NEON soil-moisture data in the MESA community folder and load it into a pandas DataFrame."
• "Launch a MESA JupyterLab analysis and tell me when it is running."

The agent calls the MESA MCP servers to do the work; see Servers for what each one can do.

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