Make a Game With Cursor: Pick a Plan, Then Play
To make a game with Cursor, install the editor, open an empty folder and ask Agent to build a small browser game that it can open and look at itself. Check the plan first, because since October 2026 new paid plans bundle Grok and charge extra for other AI.
Which plan do I need?
This is what Cursor's pricing page showed on 2 November 2026.
| Plan | Per month | What matters for your game |
|---|---|---|
| Free | $0 | Limited agent and chat requests |
| Plus | $8 | Only Grok is available |
| Premium | $30 | 4x the usage of Plus, other AI billed on top at list price |
| Super | $100 | 15x the usage of Plus |
| Ultra | $200 | 50x the usage of Plus |
Allowances refresh weekly. Cursor's announcement of 12 October says people on an older Cursor plan keep their included use of other AI until they switch, so do not switch by reflex.
The real decision is simple. If you already know whose AI writes game code you like, check whether your plan covers it. If you do not know yet, Free is enough to learn whether you like working in Cursor at all.
Why let Agent look at the game?
Much of what goes wrong in a game goes wrong on screen. Cursor's docs say Agent can "control a browser to take screenshots, test applications, and verify visual changes". For games that is its best feature, and it is easy to overlook.
Take Rooftop Runner, an invented platformer where you race from roof to roof before sunset. Agent reports that jumping works, and it does. But the runner is drawn behind the chimneys, so on several roofs you cannot see yourself. Reading the code will not show you that. A screenshot shows it at once.
Agent's browser pays off even more with a plan to follow. Ask it to draw every state the runner can be in, in Mermaid, which is text that turns into a diagram at mermaid.live. Then it can visit those states one by one and photograph each. You get a list of what it really saw, and a state it cannot reach is a bug you found early.
What if Agent breaks everything?
Agent saves a checkpoint before major edits, and you can click one to roll back. Checkpoints stay on your computer and are separate from Git, the tool for keeping versions for good. Use them to back out of a bad attempt, and ask Agent to save working versions in Git.
The Cursor setup guide covers project rules. If the jump feels off, see variable jump height.
Paste this to your AI
Build a small browser platformer called Rooftop Runner in this folder. I jump from roof to roof with the arrow keys and race the sunset to a door at the far end. Missing a roof sends me back to the last chimney.
First write a Mermaid diagram of the runner's states: running, jumping, falling, landed, back at the chimney, at the door. Check that each can be left again and that no two can be true at once. Show it to me, then build from it.
Next, use your browser tool to start the game and reach each state on the diagram. For each one, attach a screenshot and say in a sentence what is on screen, including the player, the roofs and the score. If a state looks wrong, fix that first.
If you could not open the game or reach a state, say exactly that. Your report rests on what you saw, not on what the code is meant to do.
Or skip this and connect the MCP: your AI then finds all of this by itself.Connect the MCP
How to check it worked
- Compare screenshots with your own window. They should match. If they differ, Agent looked at an old version.
- Reach every state by hand. A state on the diagram that you can never reach needs an explanation.
- Miss a roof on purpose. You should land back at the last chimney, just as the diagram says.
- Roll back one checkpoint. Do it once while nothing is at stake, so you know how.
- Look at your usage after an evening. Choose your plan from that number and not from a guess.