CLARIPHY — AI agents and Large Language Models for Scientific Discovery

Setup

Teaching: 0 min · Exercises: 0 min · Total: 0 min

Before the session

Setup is deliberately minimal — there is no Kubernetes tooling to install, no kubectl, no grid certificate. Just two things need to happen ahead of time:

1. Log into nrp.ai

Go to nrp.ai and sign in with your institutional credentials through Authentik. Access is open to users at US academic institutions and their collaborators.

If anything is unclear, Getting Started with NRP is the authoritative walkthrough — accounts, namespaces, and the portal.

2. Ask to be added to the CLARIPHY namespace

Important

An NRP account by itself is not enough to use the managed LLM service. Your account must belong to a namespace that has LLM access enabled — a per-namespace feature flag, not an account-level one.

The CLARIPHY namespace already has LLM access enabled. Message Daniel Diaz to be added to it before the session. You can see which namespaces you belong to at https://nrp.ai/namespaces/.

If you already work in another namespace that has LLM access enabled, that works just as well — you do not need to switch.

That's it for advance setup. Your personal LLM API token is the third thing you need, but it takes about thirty seconds to get and is covered — with a verification step — at the start of Introduction — LLMs on NRP. Leave it until the session starts.

How to follow along

The tutorial is built around the NRP training hub: everyone gets the same JupyterLab environment, with the materials and Python packages already installed, so nothing depends on what is on your laptop. The one exception is the JFC agent exercise, which runs on your own machine through a setup script.

A JupyterHub environment with the training materials and Python packages already installed. Nothing to install on your laptop; you just need the account, namespace, and token above.

Launch the workspace in JupyterHub

▶ Launch the workspace on the NRP training hub — signs you in at jh-training.nrp-nautilus.io, pulls the tutorial workspace, and opens JupyterLab.

Your own machine (alternative)

You can also run the notebooks on your own laptop, calling NRP's managed LLM endpoint from a local Python environment.

Important

The JFC agentic-analysis exercise — which you launch right after the introduction and check on at the end — is meant to be run locally, not on the hub: the agent installs its own toolchain and downloads ~1 GB of data, which is slow and cramped inside a hub session. A setup script does the whole install for you on macOS, Linux or Windows. If you can, grab your token from nrp.ai/llmtoken and run the script before the session, so the download isn't competing for conference Wi-Fi. If it won't work on your machine, the lesson links a JupyterHub backup.

1. Get the training materials

Open a terminal in the folder where you keep your projects, then:

Bash
git clone --branch materials/clariphy --single-branch https://github.com/nrp-nautilus/nrp-training.git clariphy
cd clariphy

This creates the tutorial folder, clariphy. The next steps all run from inside it.

  • No git? Download the materials as a ZIP, unzip it, rename the nrp-training-materials-clariphy folder to clariphy, and open a terminal inside it.
  • Already have it? Run git pull from inside the tutorial folder to update.
2. Create a Python environment

You need Python 3.9 or newer. Install the tutorial's packages into a virtual environment — a .venv folder inside the tutorial folder — rather than into your system Python. Many macOS and Linux systems now refuse a plain pip install for exactly that reason, and a virtual environment can't clash with your other projects. Deleting .venv undoes it.

macOS, Linux, WSL

Bash
python3 -m venv .venv
source .venv/bin/activate
python -m pip install -r workspace/requirements.txt

Windows (PowerShell)

powershell
py -m venv .venv
.venv\Scripts\python -m pip install -r workspace\requirements.txt

If Windows says py is not recognized, use python instead.

3. Open the notebooks

From the tutorial folder:

macOS, Linux, WSL

Bash
source .venv/bin/activate
python -m jupyter lab

Windows (PowerShell)

powershell
.venv\Scripts\python -m jupyter lab

In JupyterLab, open workspace/notebooks/. Chat with LLMs and Build a Simple Agent run as they are. The Agentic Workflows notebook uses the bash kernel the hub provides; on your own machine, run the commands from its lesson page in a terminal instead (macOS, Linux or WSL).

4. Set your token

The notebooks ask you to paste your token into their first cell. For commands you run in a terminal, set it in that terminal — it lasts until you close it:

macOS, Linux, WSL

Bash
export OPENAI_API_KEY="<paste-your-token-here>"
export OPENAI_API_BASE="https://ellm.nrp-nautilus.io/v1"

Windows (PowerShell)

powershell
$env:OPENAI_API_KEY = "<paste-your-token-here>"
$env:OPENAI_API_BASE = "https://ellm.nrp-nautilus.io/v1"

Replace the placeholder with your token from nrp.ai/llmtoken. Don't add it to your shell startup files, and don't commit it anywhere.

How to read the commands on this site
  • Shell commands are bash. They run as they are in a JupyterHub terminal, and in a macOS, Linux or WSL terminal. Where Windows PowerShell needs something different, it is shown separately.
  • Commands say where they run. The tutorial folder is clariphy — ~/clariphy on the hub, wherever you put it on your own machine. Exercises create their own new folders in your home directory (~/opencode-exercise, ~/jfc-exercise), and their commands cd there first, so it doesn't matter where your terminal started.
  • Nothing edits your existing setup. No command changes your shell startup files, the configuration of tools you already use, or your system Python.
  • Re-running is safe. If a step fails partway, fix the problem and run it again.
  • Replace placeholders such as <paste-your-token-here> before you run a command.

Getting help