Unity 6 Launches Official MCP Server — Claude Code and Cursor Can Now See Inside Your Game Scene

Unity has shipped a built-in Model Context Protocol (MCP) server as part of Unity AI, which entered open beta in May 2026 for everyone running Unity 6 (6000.0) or later. The headline for developers: AI coding assistants like Claude Code and Cursor can now read and change the live Unity editor — scene objects, component values, physics settings, and console logs — not just your C# source files.

❝Want the hands-on version? See our companion guide: Connect Claude Code to Unity 6 MCP Server.

What Happened

Unity AI ships as three parts inside the editor. The AI Assistant is an in-editor chat trained on Unity's documentation and aware of the live project. The AI Gateway lets developers route requests to third-party frontier models without leaving the editor. And the MCP Server — the piece that matters most to developers who already code with an AI agent — exposes Unity's own capabilities as MCP tools to external coding agents running in their IDE.

The MCP Server is built on the open Model Context Protocol standard, the same structured, permissioned tool interface that AI clients use elsewhere. It ships in the com.unity.ai.assistant package and is configured at Edit > Project Settings > AI > Unity MCP. The connection flow is deliberately short: Unity launches a relay binary automatically at editor startup and installs it under the user's home directory; the developer opens the settings panel, expands an Integrations section, and selects Configure for their client; Unity writes the client configuration; the developer then approves the client under Pending Connections; and the client appears under Connected Clients with a set of Unity_-prefixed tools available.

Unity's official documentation names tools such as Unity_ManageScene (open, save, and inspect scenes and the scene hierarchy), Unity_ManageGameObject (create, find, modify, and read GameObjects and their components), and Unity_ReadConsole (read editor console output, including errors and warnings). The documented built-in categories span scene management, asset operations, script editing, and console access. Rather than a fixed published list, the Unity MCP panel shows every available tool with an individual enable/disable toggle, tools are discovered and registered automatically when the editor starts, and developers can register custom tools of their own. Unity's docs list Claude Code, Cursor, Windsurf, and Claude Desktop as supported clients, and multiple MCP clients can connect to the same Unity instance simultaneously.

Why It Matters

Until now, an AI assistant working on a Unity project saw the same thing every text tool sees: code. But a large class of game bugs does not live in code. A prefab that spawns in the wrong place, an enemy that ignores gravity, a collider on the wrong physics layer, a misconfigured component reference — those are serialized scene and component values the source files never mention. A code-only assistant is structurally blind to them.

By exposing the editor's scene graph through MCP, Unity lets the assistant inspect the actual project state. An agent can open the active scene, read a GameObject's components, check the console for masked exceptions, apply a fix, and then verify it — without the developer acting as a translator between the editor and the prompt. That removes the most tedious part of AI-assisted game work: hand-describing editor state in text because the tool could not see it.

There is also a noise-control story here that experienced agent users will appreciate. Every tool exposed to a model is described to it on each request, so an overstuffed tool set dilutes focus. Unity's per-tool toggle panel turns tool selection into a per-task control — narrow it to scene and console tools while debugging, switch to asset tools during an import pass — which is a practical reliability lever, not just a settings checkbox.

What To Watch

Unity AI is in open beta, which means the APIs, the exposed tool set, and the credit-based pricing can still change; teams should check Unity's official Unity AI pricing page for the current credit allotment and MCP Server access per plan rather than treating any quoted figure as fixed. The protocol itself is open and the supported client list is already broad, so the more interesting question is not whether a given IDE can connect — it is how studios will govern permissions once an AI agent can mutate a live scene that other team members depend on. Expect tool-permission policy, not raw capability, to become the real conversation.

For now the takeaway is concrete and immediate: if you build in Unity 6 and you code with Claude Code or Cursor, your assistant can finally see the game, not just the script — and the setup is a few clicks in Project Settings.

Sources