By LLM Hardware team · Methodology ·
Checked 2026-10-13. Author: LLM Hardware team. This site does not store or show Amazon prices. Check current price on Amazon.
Key fact: Cursor Cloud Agents run in isolated cloud VMs. The docs say they were formerly called Background Agents.
How we got it: cursor.com/docs/cloud-agent, read 2 October 2026. Launch surfaces listed there include desktop, web, iOS, Slack, GitHub, GitLab, Bitbucket, Linear, and the API.
Checked:
Key fact: Continue’s public site says Continue has joined Cursor and that the open-source codebase remains available. A read-only GitHub flag was not confirmed.
How we got it: continue.dev, read 2 October 2026.
Checked:
Key fact: A local coding model at 32B Q4 needs about 21.2 GB. A 7B needs about 6.2 GB.
How we got it: 32 × 0.5 × 1.2 + 2 and 7 × 0.5 × 1.2 + 2. K2-Horizon-32B and K2-Horizon-7B are the parameter counts used.
Checked:
Methodology — parameter math, quantization bytes, and the source list.
Cursor Cloud Agents run in isolated virtual machines with a cloned repo, dependencies, secrets, and network access. The docs page says the old name was Background Agents. You start them from the desktop app (the Cloud option under the agent input), from cursor.com/agents, from the iOS app, from Slack, from a GitHub or Bitbucket comment, from Linear, or from the API. They can open a pull request, record artifacts, and hand you a remote desktop on that VM. Billing on the page is API pricing for the model you pick, and you need a paid Cursor plan. None of that uses the GPU in the tower under your desk.
The environment is the product. An agent that cannot run tests is just a diff generator, which is why Cursor spends the page on snapshots, Dockerfiles, Tailscale, and secrets. There is a separate note for running tool execution on hardware you manage (“Self-Hosted Machines”). This guide does not describe how that install works; follow their doc if you need it. Hooks from your repo can run in the cloud VM. Hooks from your home directory cannot, because the VM does not have your laptop’s home folder. That single sentence is the privacy boundary: the repo you connected is in play, your local dotfiles are not.
Grok Bot inside Cursor is a different feature. It is still their cloud computer. See the Grok Bot guide and do not mix the two names.
Three tools cover the jobs Cloud Agents advertise, with worse parallelism and better file locality.
You can run one local agent loop well. You cannot honestly claim “as many as you want in parallel” on a 24 GB card. Each extra model is another footprint. One 27B or 32B server, several clients: that is the local version of parallel agents.
Formula params × 0.5 × 1.2 + 2 at Q4. Longer context is why coding agents feel hungrier than chat. Use the VRAM calculator with the context slider up before you promise a 32K repo chat on a 24 GB card.
| VRAM / unified | Coding model at Q4 | Expect |
|---|---|---|
| 6–12 GB | K2-Horizon-7B, 6.2 GB | Small edits. Not a Cloud Agent substitute. |
| 16–24 GB | Qwen3.8-27B 18.2 GB, K2-32B 21.2 GB, Muse Glimmer ~20 GB formula | The real local coding tier. 32B is tight with a long diff. |
| 48–64 GB and up | Same 32B with a long context, or a second model | Where “more than one agent” stops being a slogan. |
The coding-model guide is about which weights. This page is about where the agent process runs.
A local coding agent wants the Studio more than the mini. On the PC side, a 24 GB card is the discrete match; an RTX 4090 listing we already use is B0BG94PS2F. A 16 GB 5080-class card does not become a 32B machine because the bandwidth is newer. Fit is gigabytes. Check current price on Amazon.
Mac Studio M5 Max 36 GB on Amazon
ASIN B0HGKSQMX6. Amazon Associates link. The live price is on that page, not locked in here.
ASIN B0BG94PS2F. Amazon Associates link. The live price is on that page, not locked in here.
Cloud: Cursor Cloud Agents are the better pick for large repos and parallel tasks: the work runs in cloud VMs on frontier models, and your GPU stays free. A local 14–32B coder is noticeably weaker on big refactors.
Local: OpenHands, Aider or Continue on your own GPU are better when code cannot leave the network, you work offline, or you want a fixed hardware cost instead of usage.
Product claims on this page carry their source and an “as of” date in the Citeable facts box. Prices are Amazon.com listings on the date given next to each one, not live prices. Reviewed .
A correction or a hardware question goes to the contact page.