# Where should your agent live? A decision sheet

**AI for Research Efficiency, episode 14.** One page, three questions. Prices, plans and limits as of 24 October 2026
(they change: check each vendor's page). Not legal advice: your campus's rules and your data's rules come first.

## 1. Do your runs outlast the laptop's lid?

Long literature runs, overnight batches, a weekly digest, anything you want to steer from your phone.

- **Your code is in a GitHub repository and needs nothing else on your computer:** a cloud agent keeps working while the
  laptop sleeps. Claude Code's cloud sessions (paid plans: Pro, Max, Team, Enterprise with premium seats) and Codex Cloud
  (paid plans: Plus, Pro, Business, Enterprise) draw on your plan's limits; Google's Jules has a free tier (15 tasks in a
  rolling 24 hours, 3 at once). They work on a copy of your project in the vendor's cloud, not on the files on your
  computer, and most need the project on GitHub.
- **It needs your own files, folders or programs:** a small always-on computer. Work through the "leave it on"
  checklist (sleep, the lid, active hours and restarts, power cuts), keep the session alive (background sessions,
  Antigravity's daemon, the ChatGPT desktop app, or tmux), and reach it safely (each tool's remote control, or SSH over
  a private network). The boxes we read cost $385 to $899 (Beelink's N150 with 16 GB, a US pre-order; the Mac mini from
  $899), and about $5 to $12 a year in power at 4 to 10 W around the clock in Arkansas.

## 2. Do you need graphics cards (GPUs) now and then?

Transcribing a batch of interviews, a local model over a large corpus, embeddings, images.

- **First, what you already have:** your campus cluster. AHPCC (the Arkansas High Performance Computing Center) is
  "available to faculty, staff and students at all of the Arkansas public universities", and "Most AHPCC services are
  provided free of charge to eligible researchers"; its public GPU jobs run for up to 72 hours (A100s). Students need a
  faculty sponsor.
- **Then, nationally:** NSF ACCESS: "It costs you nothing (really!), and you don't need an NSF award." The NAIRR Pilot:
  "all project results must be open and publishable".
- **Buy only when those do not fit** (always on, or every day, or data that may not go there). A GPU machine is a large
  purchase: a 128 GB Ryzen AI Max+ 395 mini-PC was $3,499.99 on the maker's store; Tom's Hardware reported NVIDIA's
  marketplace asking $13,250 for one RTX PRO 6000 in June 2026. At 300 W around the clock, the power alone is about $361
  a year in Arkansas (13.74 cents a kWh), about $30 a month, more than a $20 plan. On cost alone, local rarely wins (our
  reading); it wins on volume and on where the data may go.

## 3. Are the data confidential?

Student records, health data, study participants' data under an IRB protocol, export-controlled work, other people's
unpublished work.

- **Your campus's approved route first.** At the University of Arkansas: "Any AI tool (free, University licensed, not
  University licensed) used on a University system or for University business must be vetted and approved by UITS."
  The approved list names chat tools; for anything else there is a tool exception request.
- **A local model only with that approval.** Running the model locally settles where the data go; it does not by itself
  make the tool approved. Health data and controlled unclassified information bring their own rules for the machine
  itself (HIPAA's Security Rule; NIST SP 800-171). NIH's ban on AI for review critiques names the technology, not where
  it runs (our reading). Even Purdue's own campus service for open models "is not authorized for restricted data,
  including data regulated by FERPA or HIPAA".
- **The pattern that keeps data local:** let a frontier agent build and test the process on a made-up file with the same
  columns; run it yourself on the real data with a local model, on an approved machine.
- **Watch for:** Ollama's `:cloud` models run on Ollama's servers; Claude Code's remote control stores the session's
  transcript on Anthropic's servers while it is connected.

## Whatever you choose: secure it

Updates on a schedule you choose; disk encryption (BitLocker on Windows Pro, Device Encryption on Windows Home,
FileVault on the Mac); nothing open to the internet (no port forwarding on the router); SSH keys, not passwords.

## Which agent can drive a local model? (as documented on 24 October 2026)

| | Local model | How | Catch |
|---|---|---|---|
| Codex | documented | `codex --oss` with Ollama or LM Studio | use a context window of at least 64k tokens (Ollama) |
| Claude Code | through a gateway setting | `ANTHROPIC_BASE_URL`, as Ollama and LM Studio document | "Remote Control is disabled when this points at a host other than `api.anthropic.com`" |
| Antigravity | not in the CLI | the CLI documents only the Gemini API; the Python SDK can "Run Antigravity agents on your machine using local models without an API key or internet connection" | the SDK, not the CLI |

Sources: the companion page lists every one, with the date it was read.