Claude Code game development, from start to finish

Give Claude Code the project and a way to run it

Claude Code can read a whole game project, edit the files and run commands, so it can actually build something, start it and look at what went wrong.

It’s Anthropic’s coding agent, and according to Anthropic’s documentation it works in a terminal, in VS Code and JetBrains, in a desktop app or on the web.

Most of those need a Claude subscription or an Anthropic Console account. We last checked these tool details in September 2026, and the links show where each one came from.

You still pick the engine, though.

With a web library like Three.js, Claude Code writes the game code and can run and check it in Chrome itself. With Godot or Unity, it writes GDScript or C# like any other code, and an MCP server lets it work the editor as well, which you add with claude mcp add (MCP docs).

• Godot. The community godot-mcp server (MIT) launches the editor, runs the project and reads its debug output. Add it with claude mcp add godot -- npx @coding-solo/godot-mcp.

• Unity. The Unity CLI replaces the MCP server Unity deprecated in its AI Assistant package. Run unity mcp configure claude to connect Claude Code to a Unity 6.0 LTS or later editor.

If you’re on an older Unity, the community MCP for Unity server (MIT) covers a whole range, from Unity 2021.3 LTS through 6.x.

Unity even publishes an official Claude Code plugin, with 29 skills for Unity 6 or later (Unity blog), and the skills guide shows how to install it. For setup and the safety rules for both engines, see MCP servers for game development.

Set up a game folder

Start with an empty folder, initialize git, open a terminal there and run claude. To install it, Anthropic’s docs list a native installer, Homebrew and WinGet, and Claude Code asks you to log in the first time you run it.

Then add a CLAUDE.md file. Claude Code reads it at the start of every session, and /init can draft one from your project (best practices). Keep it short, though, since Anthropic warns that a bloated file makes Claude ignore your actual instructions. For a web game, four lines pretty much cover 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 the Three.js manual documents, where an import map loads the library from a CDN (Three.js manual).

Larger projects install it from npm with a build tool instead. Claude Code and Three.js walks through that setup, with a pinned release, rules and browser tests.

Write a first prompt that fits one small game

Give Claude Code a small target it can check.

Anthropic’s guide says precise prompts save corrections, and for larger features it suggests letting Claude interview you and write a spec file before any code (best practices). A first prompt in that spirit could look like this.

Interview me about a small browser game, then write the answers to SPEC.md. The game runs in one page with Three.js, works with keyboard and touch, and loads in seconds with no login. It is done when I can play a 60-second round on my phone.

Once SPEC.md exists, start a fresh session to build from it. Anthropic recommends the switch because the new session’s context holds the spec without the whole interview. And if the no-login, instant-load part sounds familiar, it mirrors the Vibe Jam 2026 rules (jam page).

It’s worth settling phones and multiplayer in the spec too, and the capybara courier, for one, supports both. Before you start building, decide which controls a player needs and whether friends join the same session.

Build one mechanic per loop

Work in loops of one feature each. Anthropic’s guide splits each loop into four phases, which are explore, plan, implement and commit (best practices).

When a change touches several files, start in plan mode, so Claude reads and proposes before it edits anything. Press Shift+Tab until plan mode shows, or start with claude --permission-mode plan. Anthropic’s rule of thumb is to skip the plan when you could describe the diff in one sentence.

Add mechanics in order (movement, then hazards, then scoring, then polish), and commit whenever a loop leaves the game playable.

There’s a /rewind command that restores earlier code and conversation, but it only tracks Claude’s own file edits, not changes made through shell commands. Anthropic says it isn’t a replacement for git, either.

If you’ve corrected Claude twice on the same problem, run /clear and rewrite the prompt with what you learned. Anthropic says a clean session with a better prompt almost always does better than a long one full of corrections, which is good to keep in mind when you’re tempted to keep arguing.

Test the game like a player

Claude needs a check it can run itself, like tests, a build or a screenshot to compare. Without one, Anthropic’s guide says, “looks done” is the only signal Claude has.

For a browser game, connect Chrome with claude --chrome. Then Claude can open your local server, read console errors and take screenshots (Chrome integration), and the nice part is that you ask for the check in plain words.

Open localhost:3000, start a round, check the console for errors and save a screenshot of the result.

You’ll need Chrome, Edge or another Chromium browser, the Claude in Chrome extension and a direct Anthropic plan like Pro or Max. The integration doesn’t run in WSL, and it stays off if you signed in with an API key.

For checks you can repeat after every change, playtesting games with AI adds a window.gameState() hook and a Playwright smoke test. Then play the game yourself, on a phone if you said it should work on mobile. As Anthropic’s guide says, “If you can’t verify it, don’t ship it.”

For touch controls, screen fit and testing on a real device, see how to make a mobile game with AI.

Publish the game

A single-player browser game only needs static hosting.

itch.io takes a zip with an index.html in it, and it requires relative asset paths, because the game is hosted in a subdirectory of its HTML CDN. The zip can hold up to 1,000 files and 500 MB unpacked (itch.io docs), so a jam-sized game has plenty of room.

Cloudflare Pages works a little differently. It deploys from a Git repository and serves the game on a pages.dev address (Cloudflare docs).

A lot of the games we list play on vercel.app, pages.dev, netlify.app or github.io addresses. Whichever you pick, ask Claude Code to check every asset path before you upload.

Once it’s live, you can submit your game.

We list games that are playable, substantially made with AI, built around a gameplay loop, safe, suitable for a general audience and not duplicates, and the editorial policy has the details. If you plan to charge for it, read selling AI-made games first.

The capybara courier credits Claude Code

Vibe Jam 2026 went to A Game About Capybaras Delivering Food (results post), where a capybara on a scooter rides through an open-world city as a food courier, and its creator lists Claude Code as the tool.

That jam required at least 90% of the code to be written by AI and drew 945 entries (jam press page), and several other highly placed games credit Claude Code as well. In our catalog it’s the most-named tool, ahead of Cursor and Codex.

Counts in the table are as of November 14, 2026 and use the tool names creators gave on the jam submission form or in a project README, which we don’t verify independently.

Measure	Claude Code games	All 588 listed games
Listed games	224	588
Vibe Jam 2026 entries	174	414
Use Three.js	131	322
No engine stated	45	151
Use Phaser	11	27
Playable on mobile	129	293
Online multiplayer	88	200

Besides that delivery game, the five examples at the bottom of the page are a stealth climber, a ragdoll brawler, a dogfighter where you draw your own plane and a kaiju roguelite. Action is the most common genre among games that credit Claude Code.

All five are Vibe Jam 2026 entries, made for a jam that ran for a month.

The creator of Floppy Brawler says Claude’s very first output, on day one, inspired the project, though they now think that early version looks terrible. All of them are on the Claude Code hub, and the jam entries are in the Vibe Jam 2026 archive.

Pick a plan and switch models with /model

Claude Code comes with the paid Claude plans, and what you pay depends on the plan and the billing period (Anthropic pricing).

Plan	Price
Pro, monthly	$20 a month
Pro, annual billing	$17 a month
Max	From $100 a month

Your usage limits are shared with the Claude chat app (Claude Help Center), so chatting and coding draw on the same allowance. With an Anthropic Console account you pay API rates per token instead.

Anthropic’s model overview lists four current models, and it suggests starting with Claude Opus 5.5 for most workloads.

Model	Anthropic’s description	API price per million tokens (input / output)
Claude Fable 5.1	Demanding reasoning and long-horizon agentic work	$10 / $50
Claude Opus 5.5	Long-running agentic coding and knowledge work	$4 / $20
Claude Sonnet 5.5	The best combination of speed and intelligence	$2 / $10
Claude Haiku 4.5	The fastest model with near-frontier intelligence	$1 / $5

To switch models in a session, use /model. Aliases like opus and sonnet point to the latest model in each tier, and opusplan is a neat one that plans with Opus, then switches to Sonnet to write the code (model configuration).

As for which model creators used, it turns out the catalog doesn’t always say. The one named most often is Claude Opus 5.5, but because the model attribution is incomplete, that doesn’t tell us which model made most of these games.

Add skills for your second game

Once the first game’s out, Claude Code skills for game development covers the skills, hooks and subagents that help with the next one. For prompts to copy at each step of a build, there’s our list of 10 prompts.

If you’re still weighing tools, best AI for game development puts Claude Code next to Codex, Cursor and three other tools, and Claude Code vs Codex for game development has the two agents side by side. Codex game development runs the same workflow with OpenAI’s coding agent.

For the basics that don’t depend on any one tool, see how to make a game with AI and vibe coding a game. Or just browse the games made with Claude and the Three.js games.