Body
What you can actually do with Cornell's managed Claude Desktop: which model to
pick and what each one costs you, how Claude handles your documents and
files, and how to save the tasks you repeat.
Worth ten minutes once Claude is installed. AI at Cornell maintains fuller
documentation in Confluence — see Further reading at the bottom.
ℹ️ Preview software. Claude Desktop is in Preview from Anthropic and
changes regularly. Features, menus, and file paths described here may shift
between updates. If something looks different on your machine, check with us.
Choosing a model
All current Claude models are available through Cornell's gateway. Claude
lists them for you — you don't configure anything, and you don't need to
remember model names. Whichever model Claude starts you on is a sensible
default; switch when you have a reason to.
There are three main models. They are the same assistant with a different
balance of how hard it thinks, how fast it answers, and what it costs
your key.
|
Thinking power |
Speed |
Relative cost |
Reach for it when… |
| Opus |
●●●●● Highest |
●●○○○ Deliberate |
$$$$$ |
The problem is hard, long, or high-stakes |
| Sonnet |
●●●●○ High |
●●●●○ Quick |
$$$ |
Almost anything — the everyday choice |
| Haiku |
●●●○○ Solid |
●●●●● Instant |
$ |
The job is short, simple, or repeated a lot |
In the model list each name carries a version number after it — Opus 5,
Sonnet 5, and so on. The number is the generation; a higher one is newer. The
family name is the part that tells you what to expect.
Opus — the heavyweight
Use it when the task has real depth: reworking a grant narrative against a
solicitation, reading a full manuscript for gaps in the argument, planning a
multi-stage analysis, or anything where being wrong is expensive.
The trade-off: it takes visibly longer to answer, and it is the most
expensive thing you can point at your key. Asking Opus for a synonym is like
booking a lecture hall for a phone call.
Sonnet — the everyday workhorse
The right default for most of the day: drafting and editing, summarising papers
and meetings, building a deck or a spreadsheet, thinking a problem through with
you. Strong enough for nearly everything, fast enough not to interrupt you.
The trade-off: on the very hardest reasoning it may need a second pass where
Opus would have got there first time.
Haiku — the quick one
Best for short, bounded, repetitive jobs: reformatting a list, tidying
references, pulling the dates out of an email thread, a first-pass sort across
many items.
The trade-off: it will take a confident swing at something complicated
rather than reasoning it through, and it holds less of a very long document in
mind at once. Give it work you could check at a glance.
A rule of thumb. Start on Sonnet. Move up to Opus when you catch yourself
re-explaining or correcting it. Drop to Haiku when you're doing the same small
thing over and over.
💰 What this means for your spend cap. Cornell bills gateway usage at
Anthropic's own published rates, with no markup — so the $ column is a
real ratio rather than a rough guide. For the same piece of work, Opus costs
about five times what Haiku does, and Sonnet about three times. The
practical consequence: a long Opus session moves your monthly actual spend far
faster than the same session on Sonnet.
Exact per-model rates are on the AI at Cornell
Available Models & Pricing
page. Your key's Spend History in the portal shows what you have actually
spent — see Managing, Rotating, and Renewing Your Key.
Fable 5 — opt-in only
Fable 5 is available, but not switched on by default. Anthropic applies
different data-retention terms to it, and you need to acknowledge those
terms before it's enabled for you.
If you'd like to use it, email itrequests@business.cornell.edu and we'll
walk you through the acknowledgement.
Working with Claude
Working with documents
Claude handles common document formats out of the box, with nothing to set up:
- Word, PowerPoint, and Excel — create or edit
- PDFs — read, summarise, or extract from
- Web pages — generate HTML and CSS
Just ask. "Turn these notes into a slide deck" or "summarise this PDF" is
enough; Claude picks the right capability itself.
Your workspace folder
Claude keeps your projects, conversation context, skills, and settings here:
|
Workspace folder |
| Mac |
~/Documents/Claude/ |
| Windows |
%USERPROFILE%\Claude\ |
Make subfolders inside it to organise projects. Claude can read anything in this
workspace when you attach it or ask for it.
Why not Documents on Windows? Cornell-managed PCs may redirect
Documents to OneDrive — whether yours does depends on how the machine is
configured. Where that redirection is in place, keeping the workspace out of
Documents avoids burning cloud storage and slowing your syncs. To check your
own machine, open Documents in File Explorer and look for a OneDrive cloud
icon in the status column.
Skills, connectors, and capabilities
These three words get used interchangeably and mean quite different things.
Knowing which one you're looking at saves a lot of confusion:
|
What it is |
Who sets it up |
| Capability |
Something Claude can already do, like building a spreadsheet from your figures |
Built in |
| Connector |
A bridge to a system Claude can't otherwise reach, like Cornell's web search |
AI at Cornell |
| Skill |
Your saved instructions for a task you repeat, like your weekly research update |
You |
The short version: a capability is what Claude can do, a connector is what
Claude can reach, and a skill is how you want it done. A skill adds no new
power — it adds consistency.
Capabilities
Capabilities are built into Claude and need nothing from you. Reading and
writing documents, building spreadsheets and decks, generating web pages,
working through calculations — all of that is already there, which is why
Working with documents above is just "ask for it."
Web search is the capability faculty ask about most. Claude searches through
Cornell's own enterprise search, and it happens automatically — when a
question would benefit from current information, Claude searches on your behalf.
You can also ask outright: "Search the web for the latest NIH grant
deadlines." The first time it searches in a session you may see a one-time
approval prompt; click Allow.
What it can reach: publicly accessible web content, current information
beyond Claude's training data, with cited source links.
What it cannot reach: Cornell systems, library databases, or anything behind
a login — and it cannot search your own files or email.
Connectors
A connector — sometimes called an MCP server — is a bridge between Claude
and an outside system: a ticketing queue, a database, a calendar, a code
repository. It is the mechanism that lets Claude reach something it has no
built-in access to.
Web search is a good worked example of the difference: searching is the
capability, and Cornell's connector is what points that capability at Cornell's
enterprise search rather than the open internet.
Cornell manages one connector for you — the one behind web search. The AI
at Cornell team configures it centrally, and it needs nothing from you.
There is no "add a connector" screen. Claude Desktop doesn't expose one, and
claude.ai connections — the catalogue of hosted integrations you'd browse
and switch on from a Claude account — aren't available either, because Cornell's
Claude Desktop never signs in to a Claude account at all (see
Appendix: Moving from a Personal Claude Subscription). For almost everyone, the managed
web-search connector is the whole story.
The advanced exception: local MCP servers can still be added by hand,
under Settings → Developer. Both kinds work — stdio (a program on your
own machine) and streamable HTTP. That's a developer-facing corner of the
app rather than an everyday feature, and most people will never need it. If
you're doing enough of this to care, the command line is the better home for it
— see Appendix: Claude Code (CLI) on Cornell's AI Gateway.
ℹ️ Who owns what. The managed connector belongs to the AI at Cornell
team, not JCB IT — if it misbehaves, tell us and we'll route it to them.
Anything you add yourself under Developer is yours to install, configure, and
keep working; we can point you in the right direction, but we can't
troubleshoot arbitrary third-party servers. Treat anything you connect as
something you're vouching for, since it runs with your access.
Skills
A skill is a saved instruction set for a task you repeat. It doesn't teach
Claude anything new; it saves you re-explaining how you want the job done.
Type / in the chat to see the ones you have and run one by name.
Good candidates are the tasks where you find yourself pasting the same preamble
every time:
- A recurring summary — "Summarise this meeting as decisions, owners, and
open questions, in that order, one line each."
- A house style — the tone, structure, and sign-off you use for student
emails or committee updates.
- A standard review pass — the checklist you apply to a draft abstract, a
syllabus, or a data table before it goes out.
- A formatting convention — how your citations, figure captions, or file
names should always look.
To add your own, the easy route is to paste its contents into the chat and say
"Save this as a skill called weekly-summary." Claude files it, and it appears
in your / menu straight away.
Skills live inside your workspace folder, so they travel with you:
|
Your skills folder |
| Mac |
~/Documents/Claude/.claude/skills/ |
| Windows |
%USERPROFILE%\Claude\.claude\skills\ |
The Code tab
Whether you have the Code tab depends on your account rather than on the
pilot as a whole. It is off by default and enabled per person on request, so
some people here have it and some don't — it follows what your work needs and
how far you want to go, not a fixed rule.
It is also a bigger step than anything else in this article: Code doesn't just
answer you, it works on your machine. That is worth understanding before you
switch it on rather than after — Privacy — What We Turned Off and Why covers what it
does, what to weigh, and how to ask for it.
Further reading
AI at Cornell publishes detailed capability documentation in Confluence,
covering all of the above in more depth. All of these ask for a NetID
sign-in.
Those pages describe the Cornell-wide deployment. Where they differ from what
you see on a JCB-managed machine, this KB is the one to trust — and tell us, so
we can get it straightened out.
Questions?
For anything about Cornell's Claude Desktop pilot — access, keys, features, or
problems — email itrequests@business.cornell.edu. We'll route it on if it
needs to go further.