Unity MCP 服务器

Watch it build a playable prototype

Neon brick breaker: scene construction, materials, gameplay scripts and visual iteration, with the assistant and Unity visible together.

▶ Open video · 25 seconds · Download MP4 · Village and castle demos

Recorded, accelerated excerpts. The silent videos export the same recordings as the GIFs; their duration is not a build-time benchmark. Media details and example prompts.

2. Install the server

Use Node.js 18 or newer; a maintained Node.js LTS is recommended.

git clone --branch main https://github.com/AnkleBreaker-Studio/unity-mcp-server.git
cd unity-mcp-server
npm ci

From environments to playable levels

Built for demanding Unity workflows

The combination is the strength: routing, scheduling, multiplayer controls and observable results in one workflow. See choosing a Unity integration for a sourced comparison.

Why choose AnkleBreaker? Each agent keeps its project identity, each editor has its own queue, and supported actions leave a history you can inspect and undo. Advanced tools extend that workflow into multiplayer iteration, Unity Hub and project-specific integrations. Those behaviors have dedicated regression checks and real-editor evidence, not just a feature list.

1. Add the Unity plugin

In Unity, open Window → Package Manager → Add package from git URL:

https://github.com/AnkleBreaker-Studio/unity-mcp-plugin.git#main

The MCP status dropdown appears automatically in Unity's main toolbar. Open the Dashboard from it, Window → AB Unity MCP, or Tools → AnkleBreaker → Unity MCP. Verify the bridge is running there. If you previously hid the indicator, choose Show Toolbar Status. Toolbar and menu access.

4. Try a complete workflow

List my running Unity projects. Select my prototype, inspect its scene and compilation errors, create a platform, then capture the scene so I can review it.

Find the multiplayer tools. List my MPPM scenarios and the players configured in the selected project.

Inspect the scene's memory use and largest objects, then show me the actions performed by each agent.

Workflows to try next

These are example prompts, not automatic scripts. Select the intended project first and let the assistant discover the current tool schemas and available packages.

Monitor and troubleshoot

<p align="center"> <img src="docs/media/dashboard-validation.png" alt="Actual Unity Dashboard in a validation project showing bridge status, three agent sessions, request queue and HTTP counters" width="640" /> </p>

Actual Dashboard capture from a Unity 6.6 validation project. The deliberately long agent/request names exercise layout behavior; displayed counters belong to that test session.

The plugin Dashboard shows queue activity, agent sessions, HTTP activity and recent actions. Agent cards update in place, sections remember their state, and long request text stays readable through tooltips. HTTP counters include input refused before ticket creation, body traffic and domain reloads. Dashboard behavior · HTTP fields and measurements.

Requested editor-window captures preserve keyboard focus, identify ambiguous windows explicitly and bound native image allocations. Docked captures exclude tab chrome; inactive tabs require an explicit temporary-switch option. Owned Windows fixtures verify actual UI Toolkit/IMGUI pixels and Dashboard layouts at 360 and 640 px.

Code execution reuses bounded compiler metadata and reports cache activity in unity_editor_state.codeExecution. A local Unity 6.6 benchmark of twenty small calls fell from 15.85 s to 1.01 s; this measures one workload, not every Unity operation. Execution behavior, measurements and limits.

• No editor found: check the plugin dashboard, console and discovered project path. See connection troubleshooting.
• Slow calls: inspect unity_queue_info, unity_agents_list and unity_agent_log. Ticket status separates queue wait and processing time; queue info also exposes inactive-session retention and evictions. Monitoring fields and measurements →
• Compilation in progress: wait for isCompiling: false, then inspect compilation errors. An empty error list during compilation is inconclusive.
• Tool unavailable: check category switches and optional packages in the dashboard. Advanced route discovery supports newer plugins without expanding the initial tool list.
• Response too large: request smaller read results; inspect the effects before repeating a write. Response budgets and image handling.
• Stopped a request: the server stops waiting and polling; accepted Unity work can still finish. Cancellation behavior.
• Client registry limit: set UNITY_MCP_COMPACT_TOOLS=1. Full parameter documentation remains available through advanced discovery.

Frequently asked questions

Which AI clients can use it? Clients that support spawning an MCP stdio server can connect using the Node command above. Configuration syntax and image/tool support depend on the client; no specific model is required by the bridge.

Do I need every optional Unity package? No. The core bridge works independently of optional integrations. Enable the relevant categories and install the packages needed for your workflow; use advanced discovery to inspect available tools.

Can several assistants work on the same project? Yes. Requests are attributed and scheduled per agent. Assign ownership of shared objects: scheduling does not resolve conflicting design changes.

Can I undo every AI action? Supported Unity writes have named undo groups and history. Filesystem operations, arbitrary code and some deferred workflows have different recovery limits. Inspect the Undo contract before relying on rollback.

Is this already a released upgrade? No. This README documents the modernization branch. Existing server/plugin version numbers remain independent; the installation instructions above select the development preview explicitly.

How does it compare? Start with the sourced comparison: routing, fair scheduling, retry recovery, multiplayer controls, monitoring and tool discovery. The measured gains above compare AnkleBreaker revisions under specific workloads.