Cursor game development in short, tested loops

Cursor’s Agent can write, run and test a browser game from your prompts, right inside an AI code editor. You still pick the library, like Three.js, and the Agent writes the game code that uses it. (If you came for mouse-cursor games or aim trainers, sorry, this page is about the editor.)

These details are from September 2026, and most of them link straight to Cursor’s docs.

Give the Agent a project folder

Cursor calls itself “your coding agent for building ambitious software”. It comes as a desktop editor, a command-line tool and cloud agents, and you can also reach it on the web, on your phone and through Slack and GitHub (Cursor).

For a game, you open your project folder in the editor and give the Agent tasks. It can search and edit files, run terminal commands, search the web, control a browser and even generate images (Agent overview, Agent tools).

You choose the model too, either Cursor’s own Composer or one from OpenAI, Anthropic, Google or SpaceXAI (models).

Unity and Godot work as well. Cursor edits the C# or GDScript, and MCP servers listed in .cursor/mcp.json let the Agent drive the editor (MCP docs).

For Unity, unity mcp configure cursor connects Cursor to Unity 6.0 LTS or later through the Unity CLI, and MCP for Unity covers 2021.3 LTS and up. For Godot, the community godot-mcp server runs the project and reads its debug output. You’ll find the setup in MCP servers for game development.

Set up a game project in Cursor

Download Cursor for macOS, Windows or Linux, sign in and pick your game folder (quickstart). Then give the Agent some standing instructions, so you don’t have to repeat yourself in every chat.

Project rules live in .cursor/rules as .mdc files, and a plain AGENTS.md at the project root works too (rules). Plain .md files inside .cursor/rules get ignored without the rule front matter, though.

Cursor’s docs say to keep rules under 500 lines. Typing /create-rule in Agent drafts one for you, and if you set it to Always Apply, it joins every chat.

For a game, a four-line AGENTS.md covers most of it.

- Stack: Three.js from a CDN import map, one index.html, modules in src/
- Run: npx serve . then open http://localhost:3000
- Controls: keyboard and touch, every action on both
- Rules: relative asset paths only; the game stays playable after every change

That’s the no-build setup from Three.js’s manual, where an import map loads the library from a CDN, while projects that use a build tool install it from npm instead (Three.js manual).

Write a first prompt that fits one small game

Start in Plan Mode, so nothing gets written before you’ve agreed on a plan. Press Shift+Tab in the chat input to switch to it, and the Agent asks clarifying questions, researches your code and drafts a plan you can edit before any code is written (Plan Mode). Your first prompt could go something like this.

Plan a one-page browser game with Three.js. Ask me questions first. The player steers a small ship with drag or arrow keys and dodges falling rocks. It works on a phone, loads in seconds and each run ends on a score screen. The plan is done when I can play one 60-second run.

Plans save to your home directory by default, but for a game project, Save to workspace keeps the plan next to the game.

Settle phones and multiplayer in the plan too. Name the controls and how another player joins, then test those choices while the game is still small.

Build in short loops with checkpoints

Add one mechanic per loop, in order, starting with movement and then hazards, scoring and polish. Cursor’s Agent creates a checkpoint before significant changes, and you can restore one from the chat timeline (checkpoints).

Checkpoints are stored locally and separately from Git, though, and Cursor’s docs say to use Git for permanent version control. So commit whenever a loop leaves the game playable.

For a bug you can reproduce but can’t explain, like a timing or performance problem, Debug Mode is pretty neat. The Agent adds log statements that report to a local debug server, asks you to reproduce the bug and reads the logs before it proposes a fix (Debug Mode).

Test the game like a player

Cursor’s Agent can drive a browser on its own. It navigates, clicks, types, scrolls, takes screenshots and reads console output and network traffic (browser tool).

Browser actions ask for your approval by default, and Cursor warns never to use auto-run mode with untrusted code or unfamiliar websites.

So ask the Agent to open the game, play one round, capture the score screen and list any console errors. For checks you can rerun after every change, playtesting games with AI adds a game-state hook and a Playwright smoke test. Then play the game yourself, on a phone if you said it should work there.

Make art with the Agent

The Agent can generate images from a text description or a reference image, and it saves them to your project’s assets/ folder by default. Cursor’s docs name UI mockups and product assets as uses (Agent tools), which isn’t really the same thing as game art.

So try it for menu art or a background, and check each image in the running game.

Some games actually skip image files altogether. According to its entry, HALDANE-4, made with Cursor, is told in ASCII characters, with procedural sound and one AI-generated voice-over.

Publish the game

A single-player browser game only needs static hosting. itch.io takes a zip with an index.html, and it needs relative asset paths, because the game runs in a subdirectory of its CDN (itch.io docs).

Cloudflare Pages deploys from a Git repository to a pages.dev address (Cloudflare docs), and each new pull request even gets its own preview URL (preview docs).

A lot of Cursor games play on vercel.app addresses, and others use netlify.app, web.app, github.io or their own domains. Before any upload, have the Agent check every asset path.

Then submit your game. The editorial policy lists what we accept, and every game is reviewed before it’s listed. If you plan to charge for it, read selling AI-made games first.

A cockpit dogfight and ASCII horror credit Cursor

The jam’s results post names Null Range for Cursor’s Best Art Direction award and HALDANE-4 for its Most Original award. Null Range, by taylor_sntx, puts you in a cockpit for first-person space dogfights online.

HALDANE-4, by denisbondare, tells a single-player horror story in ASCII characters instead. Both came out of a jam that required at least 90% of the code to be written by AI.

In the catalog’s tool counts Cursor comes after Claude Code, next to Codex. The table shows counts as of November 14, 2026, built from the tools creators named in the jam submission form or a project README, which we haven’t independently verified.

Measure	Cursor games	All 588 listed games
Listed games	106	588
Vibe Jam 2026 entries	97	414
Use Three.js	76	322
No engine stated	17	151
Playable on mobile	67	293
Online multiplayer	50	200

Cursor was the jam’s Diamond sponsor, the jam’s own page calls it the Cursor Vibe Jam 2026 (jam page), and the press page suggests Cursor and Bolt.new as tools, which may push its count up.

The counts show what creators of listed games used, not a ranking of the tools.

Cursor sometimes works next to other tools, including the Claude app. Ship Snatch lists Cursor with the Gemini app, for example, while Roomba Wars, which has a pretty great name, is a Three.js action shooter built with Cursor alone.

Action is the most common genre, though there’s an unusual one in Destined Heroes Tactics, a turn-based tactical RPG made in Unity. You can browse them all on the Cursor hub or in the Vibe Jam 2026 archive.

Start on Hobby or choose a larger allowance

You can start on Cursor’s free Hobby plan, which has limited usage. It covers Agent, Chat and Tab completions with the Auto model (pricing), which is kind of nice, because it picks a model for each request so you don’t have to (models).

Plan	Price
Hobby	Free
Pro	$20 a month
Pro+	$60 a month
Ultra	$200 a month
Teams	From $40 per user a month

Cursor’s models page lists Composer 2.5, Cursor’s own model, next to models from OpenAI, Anthropic, Google and SpaceXAI. Some creators name the model they used, but a lot of them stop at the tool.

Entries in our catalog name Claude 3.7 Sonnet and Grok 3, but with attribution this incomplete, we can’t compare the models one by one.

Connect your engine, or compare the agents

MCP servers for game development gives Cursor a browser, Blender or your engine. If you’re weighing Cursor against the agents, Claude Code vs Codex for game development and best AI for game development cover its agent-side rivals. Claude Code game development and Codex game development run the same workflow with each of them.

At Google and OpenAI, make games with Gemini and make games with ChatGPT cover the chat and agent routes.

For the tool-neutral basics, see how to make a game with AI and vibe coding a game. Or browse the Cursor games, the Three.js games and the Vibe Jam 2026 archive.