AI coding assistants for game development compared (2026)
In late 2026 the choice of assistant matters less than whether it can reach your engine. Unity, Unreal and Flutter now publish official connectors that let a general assistant such as Claude Code, Codex, Cursor or GitHub Copilot work with the project and the tooling, while Godot relies on community plugins. Entry paid plans sit between US$8 and US$20 a month, and a local model through Ollama costs nothing but hardware.
By Futigo · Updated · 8 min read
Key points
- Entry paid plans run from US$8 to US$20 a month: GitHub Copilot Pro is US$10, Claude Pro, ChatGPT Plus and Devin Pro are US$20, and Cursor Plus is US$8 but limited to Grok models.
- Unity, Unreal and Flutter each ship an official way to connect outside assistants. Unreal's is marked Experimental.
- Godot has no official AI feature that we could find. Assistants reach it through its language server and through community Model Context Protocol plugins.
- Two well-known names have changed: Windsurf now redirects to Devin Desktop, and the Gemini CLI documentation says it was replaced by the Antigravity CLI in June 2026.
- Terms differ on training: GitHub's individual terms allow training on your prompts unless you opt out, and Cursor's say it will not train on your content unless you agree.
The assistants side by side
| Assistant | Kind | Free tier | Entry paid plan | Runs offline |
|---|---|---|---|---|
| GitHub Copilot | Editor | Yes: 2,000 completions a month, limited chat and agent use | Pro, US$10 a month | No |
| Cursor | Editor, with a CLI | Yes: limited agent and chat requests | Plus, US$8 a month (Grok models only); Premium, US$30 a month | No |
| Devin Desktop (formerly Windsurf) | Editor and agent manager | Yes: light agent quota, unlimited Tab completions | Pro, US$20 a month | No |
| Claude Code | Terminal agent | No: not included in the Free plan | Claude Pro, US$20 a month, or US$17 a month billed annually | No |
| Codex | Terminal agent, with a desktop app | Yes: included in ChatGPT Free for quick tasks | ChatGPT Plus, US$20 a month | No |
| Google Antigravity | Editor and CLI | Yes: US$0 plan with weekly rate limits | Google AI Pro (price not shown on the Antigravity page) | No |
| Unity in-editor assistant | Engine built-in | Free trial on Unity Personal | US$10 a month on Personal; included with Pro, Enterprise and Industry | No |
| Unreal MCP and the Unreal Assistant panel | Engine built-in connector | Part of the editor; you bring your own agent | None listed | Depends on the agent you connect |
| Continue or Cline with Ollama | Local, open source | Yes: the software is free | None needed | Yes, with a local model |
Prices are in US dollars, before tax, from each vendor's own pricing page on 15 November 2026. Usage limits on paid plans change often, so check the linked pages before buying. The offline column for hosted tools reflects that their plans are sold as cloud services.
Some names need explaining. Cursor's pricing page now bundles Grok and X features into the same subscription, and its US$8 Plus plan offers Grok models only, with other vendors' models billed by usage from the US$30 Premium plan up. The old windsurf.com pricing address redirects to Devin's pricing page. The Gemini CLI documentation says the tool was replaced by the Antigravity CLI on 18 June 2026.
How well does each engine connect?
An assistant that only sees text files will write plausible code and miss half the project, because scenes, nodes and assets live in the editor. The Model Context Protocol (MCP) is the common way to close that gap: a small server exposes editor actions, and any compatible assistant can call them.
| Engine and language | Official AI tooling | Editor support for the language | Connectors to a running editor |
|---|---|---|---|
| Godot, GDScript | None found | Built-in language server (port 6005) and debug adapter (port 6006); official Godot Tools extension for VS Code, MIT, version 2.7.1 | Community only: Coding-Solo/godot-mcp (MIT, 6,660 stars, last commit April 2026) and hi-godot/godot-ai (MIT, 3,547 stars, version 4.4.1, needs Godot 4.7 or later) |
| Unity, C# | In-editor assistant; free Unity CLI with an MCP mode; official plugin of Unity skills for Claude Code, Codex, Grok and Antigravity | C# is a mainstream language in every editor listed above | Official MCP server through the Unity CLI (Unity 6 or later); community CoplayDev/unity-mcp (MIT, 15,982 stars, version 10.3.0) |
| Unreal, C++ and Blueprints | Unreal MCP plugin and the Unreal Assistant panel, both described in the Unreal Engine 5.9 documentation, marked Experimental | C++ is a mainstream language in every editor listed above | Official Unreal MCP server inside the editor; community ChiR24/Unreal_mcp (MIT, 1,038 stars) |
| Flutter, Dart | Official Dart and Flutter MCP server, plus a plugin with skills for Claude Code, Cursor and Codex | Dart tooling comes with the SDK | The MCP server gives analyzer diagnostics, symbol resolution, test runners and runtime inspection |
Official tooling is from the Unity, Epic, Godot and Flutter documentation. Stars, licences and versions are from the GitHub API. All checked on 15 November 2026.
Unity
Unity has the most complete official offer. Its AI tools page says the company is retiring the Unity AI brand name and now describes each tool by function: an in-editor assistant billed in AI credits, an AI gateway for agents you already pay for, an MCP server, a plugin and a CLI. The CLI, MCP server and plugin are free. Unity Muse is listed as deprecated.
Unreal
Epic's Unreal MCP documentation describes a server inside the editor with tools for actors, Blueprints, materials, Niagara, Sequencer and automation tests. It names Claude Code, Codex, Cursor, Gemini CLI and VS Code with GitHub Copilot as clients, and an Unreal Assistant panel that runs such agents inside the editor. Epic marks it Experimental and warns that the server has no authentication layer, so keep it on the local machine. Epic also runs a web-based Epic Developer Assistant, which a forum post from September 2025 said answers documentation questions and generates C++.
Godot and Flutter
Godot's documentation covers external editors through a language server, which gives an assistant in VS Code real completions and errors for GDScript while the Godot editor is running. Everything beyond that is community work, so check a plugin's last commit before relying on it. The Godot project has also tightened its own contribution rules: code sent to the engine must be written by a human. That rule covers contributions to Godot, not the games you build with it. Flutter's MCP server is official and is started by the Dart SDK's own mcp-server command.
What assistants do well in game code, and where they go wrong
They are good at the parts of a game that look like ordinary software: save files, state machines, inventory logic, grid and pathfinding maths, menus, data tables and unit tests. They are also quick at explaining an unfamiliar API and at mechanical refactoring.
They are weaker where games differ from other software. We have not seen a public benchmark for engine code that we could verify, so treat these as general observations, not measurements.
- Stale engine APIs. Engines rename and remove functions between versions. An assistant has seen more old code than new, so it can write a call that no longer exists. Less common languages such as GDScript have less public code to learn from than C# or C++.
- Scene and asset files. Godot scenes, Unity prefabs and Unreal assets carry identifiers and references that must stay consistent. Hand-edited text can break them. Editor connectors exist partly to avoid this.
- Shaders. A shader can compile and still look wrong. Only a person or a screenshot can judge the result.
- Feel. Jump height, camera lag and difficulty are tuned by playing, not by reading code.
One hands-on test of Unreal's MCP server by Puget Systems in July 2026 reported long waits per request, a change that left the test game unplayable, and heavy token use. It is one tester's account, but it matches the Experimental label.
Privacy and who owns the code
Each hosted tool sends your prompts, and the code around them, to its servers. The individual plans differ on what happens next. Business plans usually have separate terms.
- GitHub Copilot. Section J of the GitHub Terms of Service, effective 27 April 2026, says GitHub does not claim ownership of your input or output. On individual licences it grants GitHub a licence to use inputs and outputs to train models unless you opt out in account settings, and the opt-out is not retroactive.
- Cursor. The Terms of Service, updated 3 September 2026, assign to you whatever rights the company has in suggestions, and say it will not train on your content unless you have explicitly agreed.
- Claude Code on a consumer plan. Anthropic's Consumer Terms assign to you its rights in outputs, and allow training on your material unless you opt out in account settings.
- Unity. Unity says data is used only to provide the service by default and is not used for training unless you opt in.
- Codex. OpenAI's terms page would not load for us on the day we checked, so read it yourself before using Codex on private code.
None of these terms promise that generated code is free of other people's copyright. GitHub's terms say plainly that you are responsible for reviewing and testing output and for any licence it may need.
Give it context, then test what it writes
Most wrong answers come from missing context, and that is cheap to fix.
- State the exact engine version in a project instructions file that the assistant reads at the start of every session.
- Describe the project layout, naming rules and the patterns you use, such as how scenes talk to each other.
- Point it at the official documentation for your version when it gets an API wrong.
- Connect the language server or an MCP server so that it sees real errors instead of guessing.
- Ask for small changes. A short diff is easy to review, and a large one hides mistakes.
Then test. Keep game rules in plain functions that can be unit tested without the editor. Run the game after every change and read the engine's error log, because much game code fails only at run time. Commit before each task so that a bad change is one command away from being undone. Unreal's connector exposes automation tests and Flutter's exposes test runners, so an assistant can run them itself.
The local and open-source route
If your code must not leave your machine, run the model yourself. Ollama (MIT, about 187,000 GitHub stars) serves models locally. Continue (Apache 2.0) is an open-source assistant for VS Code and JetBrains editors with a CLI, and its Ollama guide describes offline workflows. Cline (Apache 2.0) also supports Ollama and LM Studio.
The cost is hardware and quality. Cline's documentation suggests 32 to 64 GB of RAM for mid-size coding models, and Continue's guide warns that not every local model handles the tool calls that agent mode needs. Expect a local model to be useful for completion and small edits, and less reliable for long agent tasks.
What we do at Futigo
Most of our games are built in Godot 4.7, and a few grid puzzles such as Parkle and Local Train Jam are plain Flutter with no game engine. Much of the art in our Godot games is drawn in code, and our sound effects and music loops are synthesised by small Python scripts kept in each game's repository. Our engine comparison is in Open-source game engines compared.
Questions
- What is the best AI coding assistant for game development in 2026?
- There is no single winner. Pick one your engine officially connects to: Unity's plugin supports Claude Code, Codex, Grok and Antigravity, and Epic's Unreal MCP documentation names Claude Code, Codex, Cursor, Gemini CLI and GitHub Copilot in VS Code.
- Is there an official AI assistant for Godot?
- We found none. Godot offers a language server for external editors and an official VS Code extension, and assistants connect to a running editor through community MCP plugins such as Coding-Solo/godot-mcp and hi-godot/godot-ai, both MIT licensed.
- How much do AI coding assistants cost in 2026?
- As of November 2026, GitHub Copilot Pro is US$10 a month, Claude Pro, ChatGPT Plus and Devin Pro are US$20 a month, and Cursor starts at US$8 a month for Grok models only. Copilot, Cursor, Codex, Devin and Antigravity have free tiers.
- Does Unreal Engine have built-in AI coding tools?
- Yes, in Experimental form. Epic's Unreal Engine 5.9 documentation describes an Unreal MCP server inside the editor and an Unreal Assistant panel that runs agents such as Claude Code, Codex, Cursor and OpenCode.
- Can I use an AI coding assistant offline for game development?
- Yes, with a local model. Ollama runs models on your own machine, and the open-source assistants Continue and Cline can use it. Cline's documentation suggests 32 to 64 GB of RAM for mid-size coding models.
- Do I own the code an AI assistant writes for my game?
- The terms we read from GitHub, Cursor and Anthropic do not claim ownership of output, and Cursor and Anthropic assign their rights in it to you. None guarantee that the output is free of third-party copyright, so review it as you would any outside code.
Sources
Checked on 15 November 2026.
- GitHub Copilot plans and pricing
- Cursor pricing
- Claude pricing (plans that include Claude Code)
- Codex pricing
- Devin pricing (windsurf.com redirects here)
- Google Antigravity pricing
- Gemini CLI quota and pricing
- Unity's AI tools
- Unity's plugin for third-party agents
- Unreal MCP in Unreal Editor (Epic documentation)
- Godot documentation: using an external text editor
- Godot: changes to our contribution policies
- Dart and Flutter MCP server
- GitHub Terms of Service (Section J)
- Cursor Terms of Service
- Anthropic Consumer Terms of Service
More insights
- Mobile game retention benchmarks (2026)Games industry
- What it costs to make a small mobile game (2026)Games industry
- Localising a mobile game into Indian languages (2026)Localisation
- Open-source 3D tools for game development (2026)Game art
- Game engine pricing compared (2026)Game development
- Godot 4.7 for mobile games: what changed and what to set up (2026)Game development
We make small games and apps. Here are a few.
Futigo on Google Play- Ants Flood EscapeDig channels to steer the flood away and keep every ant dry.
- ClaywheelShape a pot on the wheel, paint it and fire it in a glowing kiln.
- ParkleTap the cars out of a packed parking lot in the right order.
- Local Train JamClear the crowd in a local train and get off at your station.
- Power CutThe lights are out: grab the torch and find what the family needs.
- 10-Minute RiderWeave a delivery scooter through rainy traffic against the clock.