Draft for review. This page is under review before the Operational Readiness Review. Panels marked as open questions are still waiting for an answer. Send corrections to the Research Platform team.
Read time: 6 minutes. Short version: JupyterLab has a built-in AI assistant. You choose which AI model it talks to. You can bring your own, buy a GitHub Copilot seat through the Science Cloud, or ask to join the pilot for the Science Cloud's own model.
What Notebook Intelligence is
Notebook Intelligence is the AI assistant built into your workspace. It is not a separate web page. It sits next to your notebook.
An LLM (large language model) is the AI model that answers. Notebook Intelligence is the part in JupyterLab; the LLM is the brain it talks to. You choose the brain.
What it can do
| Feature | What it means for you |
|---|---|
| Chat | Ask a question in a side panel and get an answer about your code or your data |
| Inline completion | It suggests the rest of the line or block as you type |
| Agent mode | It can create a notebook, edit cells, and run them for you |
| Tools (MCP) | It can use extra tools you give it, such as a data-catalog search |
| Approval step | For actions that change things, it asks you first |
Examples
Explain. Select a cell you do not understand and ask what it does.
Write. In an empty cell, type "read this CSV, drop empty rows, plot the monthly mean". Read the code it writes. Then run it.
Fix. A cell fails. Ask what the error means and let it correct the cell.
Do a job. In agent mode, ask it to build a small notebook that loads a file and makes three charts. It creates the cells and runs them. It asks before it changes anything.
Use a tool. In the AgentRegistry panel, press Add to NBI on an approved MCP server. The assistant can then use that tool in your next question. See the tool reference.
Your three ways to get a model
| Way | What you do | Who pays |
|---|---|---|
| 1. Bring your own | Use your own licence or API key from a provider | You or your project, direct to the provider |
| 2. GitHub Copilot through the Science Cloud | Ask the Science Cloud for a Copilot subscription | Through the Science Cloud |
| 3. Skippy — the Science Cloud's own model | Ask to become a pilot user | The Science Cloud |
Skippy is in pilot. The Science Cloud's internal model, Skippy, is not a general service yet. It is running as a pilot. Ask support@sciencecloud.nasa.gov to be added as a pilot user, and expect pilot-level service.
Coming later. Running models through the AI Registry, and through AWS Bedrock. Not available today.
Which providers can I connect?
Notebook Intelligence supports these provider types:
| Provider | Notes |
|---|---|
| GitHub Copilot | Sign in with your GitHub account |
| OpenAI-compatible | Any service that speaks the OpenAI API — this is the setting to use for most hosted models, and for the Science Cloud's own endpoint |
| LiteLLM-compatible | For a LiteLLM gateway |
| Ollama | Models running locally, in your own workspace |
| Anthropic Claude | Through the Claude Code command-line tool, in Claude mode |
How to connect a model
- In JupyterLab, open the assistant's chat panel.
- Click the gear icon. (Or use Settings → Notebook Intelligence Settings.)
- Choose the provider type from the list above.
- Fill in what it asks for — usually a model name, an address, and an API key.
- Save. Then ask the assistant a question to check it works.
GitHub Copilot is different: you sign in to GitHub instead of pasting a key. You can tick "Remember my GitHub Copilot access token", which stores an encrypted token in ~/.jupyter/nbi/user-data.json.
Where your settings are kept
Your settings live in ~/.jupyter/nbi/config.json, inside your home folder — so they survive a restart. There can also be a shared default for the whole software set.
Your API key is stored in plain text. An API key you type into the settings is written into ~/.jupyter/nbi/config.json unencrypted. That is how the tool works, so treat that file with care: - Never commit it to Git. Not to any repository, public or private. - Never copy it into a shared folder such as /efs/<project>, and never send it to anyone. - Never put a key in a notebook cell either. - If a key does get out, cancel it at the provider immediately and create a new one. Your home folder is private to you. Keep the key there and nowhere else.
The rule about which models you may use
This part matters. Read it before you connect anything.
Do not use a model built in a country on NASA's designated country list. Models from countries that are not on that list are generally acceptable.
The list is published by NASA: <https://www.nasa.gov/wp-content/uploads/2026/08/designated-country-list-8-4-2026.pdf>
If you are not sure about a model, ask before you use it. Write to support@sciencecloud.nasa.gov. The Research Platform support team will tell you. Asking takes a few minutes. Guessing wrong is a security incident.
Two more habits worth keeping:
- Think before you send data. When you use an outside provider, your prompt leaves the platform. Do not paste data your project would not put in an email to that company.
- Check the answer. The assistant can be confident and wrong. Read the code before you run it, and check the science yourself.
Common questions
Do I need a key to start? For GitHub Copilot you sign in instead. For most other providers, yes, you need a key or a licence. For Skippy, you need to be a pilot user.
Does the platform pay for my model use? Only for the routes the Science Cloud provides — a Copilot subscription through the Science Cloud, or Skippy. If you bring your own key, you pay the provider yourself.
Can I run a model with no internet provider at all? Yes. Ollama runs a model inside your own workspace. It is limited by the size of the machine you picked, and the machine's cost still applies.
Do my settings follow me to a new workspace? Yes, if they are in your home folder. They do not follow you into a collaboration session, which is a separate shared workspace.