MCP servers for game development: what to install

An MCP server gives a coding agent tools it doesn’t have by default: a browser to play your game in, a line into Blender, or control of the Godot, Unity or Unreal editor. Install a browser server first, then one for the tool you build in.

MCP, the Model Context Protocol, is the open standard behind that connection. A server exposes tools the agent can call, resources it can read and prompt templates (MCP docs). The current specification is dated July 28, 2026 (versioning), and everything below is as of September 2026.

The short list of game-dev MCP servers

One browser server and one server for your engine or art tool is enough to start.

Server	What your agent can do with it	Official?
Playwright MCP	Open the game, click, press keys, screenshot, read the console	Yes, Microsoft
Chrome DevTools MCP	Record performance traces, throttle the CPU or network, run Lighthouse	Yes, Chrome team
Blender MCP	Run model-written Python in your Blender session	Yes, Blender Lab
mcp-for-blender	Scene control plus asset search and 3D generation services	No, community
godot-mcp and godot-ai	Launch the project, read debug output, edit scenes	No, community
Unity CLI (unity mcp)	Drive a running Unity editor	Yes, Unity
Unreal MCP	Editor toolsets in Unreal Engine 5.8	Yes, experimental
Aseprite servers	Draw and edit pixel art	No, community
Context7	Current, version-specific library docs in the prompt	Yes, Upstash

Browser servers: let the agent play your game

Playwright MCP is the one to install first for any browser game. In Claude Code:

claude mcp add playwright npx @playwright/mcp@latest

Codex takes codex mcp add playwright npx "@playwright/mcp@latest", and in Cursor you add npx @playwright/mcp@latest under Settings, then MCP (README). The agent then gets tools such as browser_navigate, browser_press_key, browser_evaluate, browser_console_messages and browser_take_screenshot.

Playwright MCP normally acts through the page’s accessibility tree, and to that tree a canvas game is a single element. Its vision capability (--caps=vision) adds coordinate tools such as browser_mouse_click_xy, and Playwright’s docs name canvas and WebGL apps as the use case (vision mode). The loop becomes screenshot, click at a coordinate, screenshot again.

Better still, let the game tell the agent what is happening. browser_evaluate runs JavaScript in the page, so a game that exposes a small state function on window (score, lives, player position) can be read directly. Playtesting games with AI builds that hook step by step.

Playwright’s README also points coding agents to its CLI with skills as the more token-efficient option, and keeps MCP for exploratory sessions (README, Playwright CLI). If your agent runs shell commands anyway, try the CLI first.

Chrome DevTools MCP observes where Playwright acts. performance_start_trace records a trace with Core Web Vitals, and emulate throttles the CPU or the network (tool reference). lighthouse_audit covers accessibility, SEO and best practices but leaves performance to the trace. Ask the agent to record ten seconds of play on a throttled CPU for a rough read on how the game runs on a cheap phone.

claude mcp add --scope user chrome-devtools -- npx -y chrome-devtools-mcp@latest

Google collects usage statistics by default; the --no-usage-statistics flag turns that off (repo).

Blender: from prompt to model

Blender has an official MCP server from Blender Lab, made of an add-on plus a standalone server, and it needs Blender 5.1 or newer (Blender Lab). Anthropic announced it in April 2026 as a connector built by the Blender developers (Anthropic). The model works through Blender’s Python API: it writes the code, and the server runs it in your open session. Blender’s examples include finding performance bottlenecks in a scene and documenting geometry node setups.

Blender’s page warns that the server runs model-written code “without any guards in place” to stop your data being deleted or sent elsewhere. It advises a virtual machine or a computer without sensitive data (Blender Lab), and that advice deserves to be taken literally.

The community option, mcp-for-blender (formerly blender-mcp, MIT), adds search and download for Poly Haven, Sketchfab and Poly Pizza, and generation through Hyper3D Rodin, Hunyuan3D and Tripo (README). It calls itself a third-party integration, not made by Blender. With BLENDER_MCP_SAFE_MODE=1 it checks every script first and blocks file access, other programs and network calls.

For a first task, ask for something with a number in it:

Make a low-poly treasure chest under 2,000 triangles and export a GLB.

For models from a picture instead, see image to 3D for games.

Game engines: Godot, Unity and Unreal

Godot has no official server. Coding-Solo/godot-mcp (MIT) adds with claude mcp add godot -- npx @coding-solo/godot-mcp and offers launch_editor, run_project, get_debug_output, create_scene and add_node (README). Versions before 0.1.1 passed the project path to a shell, so a crafted path could run commands on your computer; update (advisory).

hi-godot/godot-ai (MIT, Godot 4.7+) goes wider, with 46 tools that build scenes, edit nodes and scripts, wire signals and set up UI, materials and animation (README). satelliteoflove/godot-mcp can freeze the game clock, step exact slices of time and inject input, which is what a repeatable playtest needs (README).

Unity deprecated the MCP server inside its AI Assistant package and moved it to the Unity CLI, which drives a running editor on Unity 6.0 LTS or later (Unity docs). After installing the CLI:

unity pipeline install
unity mcp configure claude

The client name can also be cursor, vscode or windsurf, and unity mcp configure --list shows the rest. The community CoplayDev/unity-mcp (MIT) supports Unity 2021.3 LTS through 6.x with tools such as manage_scene, read_console and run_tests (README).

Unreal Engine 5.8 ships an experimental Unreal MCP plugin that serves on http://127.0.0.1:8000/mcp. The console command ModelContextProtocol.GenerateClientConfig ClaudeCode writes the client config, and toolsets such as SceneTools and ActorTools come from Epic’s toolset plugins, most of them written in Python. Epic’s docs say there is no authentication layer, so keep it on your own machine (Epic docs).

Pixel art, generated assets and docs

For sprites, diivi/aseprite-mcp (MIT) exposes over a hundred Aseprite tools, including a Lua escape hatch that runs unrestricted code on your computer (README). For image, audio and 3D models from hosted services, fal.ai runs an official remote server; the server is free and each model run is billed (fal docs).

Context7 puts up-to-date, version-specific documentation and code examples for your libraries into the prompt, so the agent writes against the version you actually installed. It is free, with an optional key for higher rate limits, and runs as an MCP server or as a CLI with skills (Context7). Codex’s docs use it as their example: codex mcp add context7 -- npx -y @upstash/context7-mcp.

Set it up once, per project

In Claude Code, --scope project writes the server to .mcp.json at the repo root, which you commit so the team gets the same tools (Claude Code MCP docs). Keep keys out of that file with ${VAR} placeholders, which Claude Code expands from your environment; ${VAR:-default} adds a fallback.

Codex reads servers from ~/.codex/config.toml, or from .codex/config.toml in a trusted project, and the ChatGPT desktop app, the CLI and the IDE extension share that setup (Codex docs). Cursor keeps project servers in .cursor/mcp.json (Cursor docs).

Enable only the two or three servers you use this week. Every server adds tool descriptions to the agent’s context, and Anthropic calls command-line tools such as gh “the most context-efficient way” to reach outside services (best practices).

Safety: three rules

The MCP specification says tools represent arbitrary code execution and that hosts must ask for consent before calling them (spec). Claude Code asks before it uses a project’s .mcp.json servers in an interactive session. In claude -p runs, Agent SDK sessions and cloud sessions it loads them without asking, unless you block them with --strict-mcp-config or disabledMcpjsonServers (Claude Code MCP docs).

1. Install from sources you trust, and pin versions. Read the README, check who maintains it, update when an advisory lands.
2. Keep the lethal trifecta apart. Simon Willison’s rule is to never give one agent private data, untrusted content and a way to send data out at the same time (Willison). A browser server reads untrusted pages, so don’t pair it with your secrets.
3. Sandbox anything that runs code. Blender, Aseprite’s Lua tool and Unreal’s unauthenticated endpoint all run what the agent writes. A separate user account or a virtual machine is cheap insurance.

Next: put the servers to work

• Claude Code skills for game development: skills and hooks that tell the agent when to use these tools.
• Playtesting games with AI: Playwright in practice, with a game state hook.
• Best game engine for vibe coding: Three.js, Phaser, Godot and Unity side by side.
• Browse the Godot games and Unity games in the catalog.