Cursor shipped a first-class Agent CLI — same agentic loop, portable binary, installable with one curl. Hosting "Cursor" as a desktop app on a VPS is the wrong shape. Hosting agent on its own always-on Firecracker microVM is the right one: login in a remote PTY, rules in the repo, MCP when you need tools, and a blast radius that ends at the VM wall.
Here's the clean setup on jurniti Cursor Agent hosting.
What Cursor Agent is
Cursor Agent is Anysphere's terminal coding agent. Official install:
curl https://cursor.com/install -fsS | bash
That drops agent (and cursor-agent) under ~/.local/bin, with the real binary versioned under ~/.local/share/cursor-agent/versions/.
Useful entry points:
agent— interactive session in the current workspaceagent login— account auth (NO_OPEN_BROWSER=1for SSH/remote)export CURSOR_API_KEY=…— automation / CIagent --print -p "…"— non-interactive one-shotagent mcp,agent models,agent status
Settings live in ~/.cursor/cli-config.json. Session auth is separate (do not publish it).
Why Cursor Agent needs its own microVM
Cursor Agent acts. It edits the workspace, can run tools, and can call MCP servers. On a laptop that's your whole home directory. A shared container still shares the host kernel.
jurniti gives Cursor Agent its own Firecracker microVM — own kernel, KVM boundary, one tenant per box. A bad tool call or a poisoned MCP integration stays disposable.
You keep the Cursor relationship. Login or CURSOR_API_KEY stays inside the guest. jurniti is not in the model path and does not mark up tokens.
Most work fits Starter (1 vCPU / 2 GiB). 30-day money-back, workspace kept after cancel per plan terms.
Cursor's cloud agents vs your own microVM
Cursor also sells cloud agents — Anysphere-hosted workers running in their infrastructure. They're good at parallel PR grinding inside Cursor's product; the tradeoff is that the workspace, the schedule, and the box all belong to Cursor. A personal microVM is the other shape: an always-on agent CLI in a box you control — your dotfiles, your cron jobs, your MCP config, persistent between sessions — with a KVM boundary instead of a multi-tenant substrate. If the job is "my agent, my workspace, always on," this is that.
Run Cursor Agent in 3 steps
1. Pick a plan and pay
Starter is enough for most Cursor Agent work. Checkout triggers auto-provision.
2. Your microVM boots with agent ready
Ubuntu guest, CLI on PATH, in-browser terminal — no SSH required:
NO_OPEN_BROWSER=1 agent login
# or: export CURSOR_API_KEY=...
cd your-project
agent
3. Put it to work
Add project rules (AGENTS.md / team conventions), wire MCP, use --print for jobs you don't want to babysit. Config survives restarts on the persistent home.
What people run Cursor Agent for
- Always-on agentic coding without pinning a laptop
- Remote PTY login with
NO_OPEN_BROWSERdevice/code flows - Headless jobs via
agent --print - MCP-heavy workflows with secrets that never leave the guest
Snapshot a configured agent into a template — settings and rules, never auth.
jurniti vs laptop vs DIY VPS
| jurniti | Laptop | DIY VPS | |
|---|---|---|---|
| Always on | Yes | No | Yes |
| Isolation for tool-using agents | Firecracker (KVM) | Whole machine | Shared kernel |
| Cursor credentials | Yours, in the VM | Yours | Yours |
| Setup | ~3 min automatic | Local install | Manual forever |
| Refund | 30-day money-back | n/a | None |
Pricing
Starter $25/mo ($250/year). Pro $49/mo, Max $99/mo. Model spend separate, never marked up. Every plan: 30-day money-back.
Running a different agent? See Claude Code, Devin, or CLI agents compared, get managed Cursor Agent hosting, or see pricing.