Unity Workflow LibraryUWLAbout

Topic 4 · Verify & test

Unity MCP tooling options

The Unity MCP servers that exist, what each one does, how to pick one, and why the library's rules work with any of them.

Markdown version

The Model Context Protocol (MCP) lets a coding agent call tools in a running Unity Editor: read the console, run tests, enter Play mode, inspect and edit scenes and assets. There is no single standard Unity MCP server. Several maintained projects exist, plus Unity's own option. This page lists the ones checked in October 2026, with their repositories, so you can pick one. Tool counts and features change quickly; read each project's README before installing.

How MCP fits the workflow

  • The loop is: write, compile, read, prove. The agent edits code or assets, lets the Editor compile, reads the console for errors, runs the relevant tests, and then proves the result with an output it can see (a log line, a test result, a screenshot, a value read back from the scene).
  • Typical calls in that loop: get console logs; get editor state (compiling, playing, paused); enter, pause, step and exit Play mode; run EditMode and PlayMode tests; list the scene hierarchy and read component values; create or change GameObjects, components, prefabs and assets; take a screenshot.
  • Two halves. Most options have an Editor package (a bridge that runs inside Unity) and a separate MCP server process your agent client launches (Node.js, Python or a binary). Both halves must be installed and must match versions.

The tools

  • AnkleBreaker Unity MCP (AnkleBreaker Studio).
  • What it does: a very broad tool set for the Editor and Unity Hub: scenes, GameObjects, components, builds, profiling, Shader Graph, Amplify Shader Editor, terrain, physics, NavMesh, animation, tests and multiplayer Play mode scenarios (MPPM, ParrelSync).
  • Install: add the Editor plugin with Package Manager, Add package from git URL https://github.com/AnkleBreaker-Studio/unity-mcp-plugin.git (the bridge listens on localhost:7890), then run the Node.js server (npx anklebreaker-unity-mcp@latest, or clone and npm install).
  • License: free and open source under the custom "AnkleBreaker Open License v1.0" (attribution required); read it before redistributing.
  • Links: unity-mcp-server · unity-mcp-plugin.
  • MCP for Unity (CoplayDev). The project formerly at justinpbarnett/unity-mcp, which now redirects here.
  • What it does: manage assets, control scenes, edit scripts, read the console, run tests and automate Editor tasks; publishes a full tool catalog.
  • Install: Package Manager git URL https://github.com/CoplayDev/unity-mcp.git?path=/MCPForUnity#main (or OpenUPM com.coplaydev.unity-mcp); the server needs Python 3.10+ through uv. Supports Unity 2021.3 LTS to 6.x.
  • License: MIT. Link: CoplayDev/unity-mcp.
  • Unity-MCP (IvanMurzak).
  • What it does: 70+ built-in tools (console logs, editor state and Play mode control, EditMode and PlayMode tests, packages, scenes and assets), custom tools from any C# method with one attribute, agent skills, and a mode that runs inside a built game for runtime debugging.
  • Install: OpenUPM com.ivanmurzak.unity.mcp or the .unitypackage installer; or the CLI (npm install -g unity-mcp-cli, then unity-mcp-cli install-plugin ./MyUnityProject).
  • License: Apache-2.0. Link: IvanMurzak/Unity-MCP.
  • MCP Unity (CoderGamester).
  • What it does: scene and GameObject tools, run_tests through the Test Runner, get_console_logs with paging, get_play_mode_status and set_play_mode_status (play, pause, stop, step), package install, plus resources and guided prompts.
  • Install: Package Manager git URL from the repository, then configure the client from Tools > MCP Unity > Server Window; needs Node.js 18+.
  • License: MIT. Link: CoderGamester/mcp-unity.
  • Unity's own option: the Unity CLI unity mcp.
  • Status: the in-Editor MCP server that shipped in the AI Assistant package (com.unity.ai.assistant) is deprecated. Unity replaced it with the MCP mode of the Unity CLI, built on the Unity Pipeline package; unity mcp configure <client> writes the client configuration. Unity also recommends agents that can run shell commands call the CLI directly instead of MCP.
  • Links: Unity CLI replaces the in-Editor MCP server · Unity CLI reference.
  • Smaller projects exist too, for example mitchchristow/unity-mcp (MIT, small). Some are archived, such as nurture-tech/unity-mcp-server. Check activity before adopting one.

Picking one

  • PICK-BY-THE-LOOP-YOU-NEED — choose the server that covers your verification loop, not the longest tool list.
  • Rule: before installing, confirm it can read console logs, report compile and Play state, run EditMode and PlayMode tests, read the scene hierarchy and component values, and take a screenshot. Add domain tools (Shader Graph, terrain, multiplayer, builds) only if you use them.
  • Why: a server missing console or test access forces the agent to guess, and guessing is how unverified changes ship.
  • Check: on a fresh project, ask the agent to make a compile error, read it from the console, fix it, and run one test. All four steps work through the server.
  • PIN-THE-VERSION — pin the Editor package and the server to matching versions.
  • Rule: install from a tagged release or a pinned git revision, and upgrade both halves together.
  • Why: these projects release often; a server and bridge on different versions fail in confusing ways.
  • Check: the project manifest and the client configuration name the same release.
  • CHECK-LICENSE-AND-ACTIVITY — read the license and recent activity before adopting.
  • Rule: MIT and Apache-2.0 are permissive; a custom license needs reading. Prefer projects with recent releases and an active issue tracker; avoid archived ones.
  • Check: the license is recorded next to the dependency, as on the Legality & liability page.
  • LOCAL-ONLY-AND-APPROVED — the bridge listens on localhost, and you know what the agent may change.
  • Rule: keep the Editor bridge bound to the local machine, approve client connections where the tool asks, and use version control so every agent write can be reviewed and reverted.
  • Check: no bridge port is reachable from another machine; agent changes appear as reviewable diffs.

Using it well

  • MCP-IS-A-TRANSPORT-NOT-A-PROOF — a tool call that returned "ok" proves nothing about the game.
  • Rule: after every write, prove the effect from its output: read the value back, read the console, run the test, or look at a screenshot.
  • Why: a call can succeed while the change targets the wrong object, the wrong Editor, or never renders.
  • Check: every change in a session has a proof line next to it.
  • READ-THE-CONSOLE-AFTER-EVERY-WRITE — compile errors and exceptions are read by the agent, not reported by the user.
  • Rule: after a script or asset write, wait for compilation, then read errors and warnings before the next step.
  • Check: the session log shows a console read after each write.
  • ONE-AGENT-ONE-EDITOR-ONE-SERVER — one agent drives one Editor through one server.
  • Rule: do not run two MCP servers against the same Editor, or two agents in the same project at once; with several Editors open, confirm which project the server is talking to before each phase of work.
  • Why: concurrent writers trigger recompiles, end Play sessions and route calls to the wrong project.
  • Check: a call that reads the project path returns the project you expect.

The library's rules work with any of them

The rules on this site are tool-agnostic. They describe how Unity, C# and platforms behave, and how to prove a change works; they hold whichever MCP server, CLI or manual workflow you use. The ones that matter most when an agent drives the Editor:

Checklist

  • The server covers console logs, compile and Play state, tests, scene reads and screenshots.
  • Editor package and server pinned to one matching release.
  • License read and recorded; the project is active, not archived.
  • Bridge is local only; agent writes go through version control.
  • Every write is followed by a console read and a proof from output.
  • One agent, one Editor, one server; the target project is confirmed per phase.