Unity’s Official Claude Code Plugin Is Not Open Source

Unity shipped its official plugin for Claude Code today. It sits in Anthropic’s managed plugin directory, which makes Unity one of the first major engine vendors to put its trademark on agent tooling. Three things the announcement skipped. The plugin runs under the Unity Companion License rather than anything open source, so GitHub refuses to classify the repo at all. The blog claims it installs Unity’s MCP server for live Editor control while the plugin’s own README says it ships no MCP servers. Then there’s the in-app purchases skill, which carries dedicated reference files for stripping RevenueCat or Adapty out of your project before replacing them with Unity IAP. The skill count is 31, not the 29 Unity announced. Version reads 0.1.2-beta. Best for Unity 6 developers already running Claude Code every day. Not ideal for anyone on older Unity versions or expecting Editor control the moment it installs.

Thirty one skills. Roughly sixty thousand words of instructions. Two files explaining how to remove a competitor’s software from your project.

Unity shipped its official Claude Code plugin this morning. The announcement mentions none of that.

What it does mention is that Unity engineers wrote these skills against the engine’s documented behaviour rather than whatever a model absorbed during training. That part is true. It’s also useful, which is why anyone building in Unity 6 should install this today.

Then you open the license file.

What Actually Installs

Every row below came out of the repository rather than the launch post. I pulled the current build then read the manifest.

Three rows there contradict Unity’s own launch material.

Nobody Mentioned the License

The repo sits on GitHub. For most developers that carries exactly one meaning: the code is open, so take it.

This code is not open.

What the Unity Companion License allows

The plugin ships under the Unity Companion License, which permits use only inside projects that depend on Unity. Build your Unity game with these skills all day. You cannot fork them into something competing with Unity. You cannot lift the patterns into an engine Unity doesn’t own.

GitHub’s license detector reads that file, fails to match it against anything recognised, then assigns nothing at all. Scan the repo page for a license badge and you find blank space where MIT usually sits.

Why the silence matters

Unity’s announcement never mentions the license. Neither does the directory listing. For a launch aimed at developers who install things from GitHub on reflex, leaving that out does real work.

To be fair, restricting reuse is defensible. Unity wrote these skills. Unity pays the engineers maintaining them. Protecting an asset that cost money to produce is ordinary business.

What’s less ordinary is the inversion. Most community Unity plugins ship under MIT. Developers read them, fork them, then carry the patterns to Godot if they feel like it. That door shuts here, quietly, on launch day.

Anyone who watched AI vendors adjust terms on people this past year recognises the shape. The restriction is rarely the problem. Announcing everything except the restriction is.

The Count Is Wrong

Unity’s launch post says 29 skills. Unity’s repository ships 31.

I counted the skill definition files in the current build. There are 31 of them. The repo also took commits within hours of the announcement, so two almost certainly landed after somebody locked the blog copy.

Nobody is being misled by this. Still, if you’re writing about the launch, use 31, because 31 is what lands on your machine.

Unity Wrote Two Files About Deleting RevenueCat

Open the in-app purchases skill, look in its reference folder, then find these: convert-revenuecat.md and convert-adapty.md.

Each runs a structured assessment for a project already using a competing monetization SDK. Each includes a section covering complete removal of that competitor before Unity IAP goes in.

How deep it runs

Across all 31 skills I counted 107 references to Adapty, 106 to ironSource, 74 to RevenueCat, 50 to UniPay then 43 to Essential Kit. Those aren’t compatibility footnotes. They sit inside numbered migration paths, complete with a required migration report the agent produces before touching any code.

Picture the sequence. A developer opens Claude Code, types something ordinary about cleaning up a purchase flow, then gets an agent already holding Unity’s framework for whether RevenueCat can come out.

The guidance itself is fair, which is stranger

Unity’s own file states that RevenueCat calls Apple StoreKit plus Google BillingClient directly through its native SDKs, so no clean swap exists. It instructs the agent to assess how much of RevenueCat’s value Unity IAP can replace before recommending anything. Several passages flag conditions that should stop a conversion outright.

That’s more candid than most vendor migration docs manage. Which is what makes the placement worth sitting with: competitive positioning now lives inside the context window, arriving before the developer has formed a preference, phrased in the voice of an assistant they already trust.

The MCP Server Isn’t In the Box

Here Unity’s marketing and Unity’s repository disagree with each other.

Two sources, one contradiction

The Unity blog says the plugin installs Unity’s engineering skills, the Unity CLI, plus Unity’s MCP server for live Editor control.

The plugin’s README says it ships skills only, with no hooks or MCP servers.

The README is correct.

What live Editor control actually takes

One skill, unity-cli, knows how to configure Unity’s MCP server on your behalf. Running that configuration needs the Unity CLI already installed. It also needs the com.unity.pipeline package added to your project through a separate command.

The capability is real once wired up. Claude can create GameObjects, edit scenes, inspect the hierarchy then execute C# against a running Editor instead of hand editing scene YAML. Every Unity developer who has watched an AI corrupt a prefab understands why that matters more than any skill in the collection.

It simply takes three installs rather than one command. Read the blog post, expect Editor control after installing the plugin, then spend twenty minutes hunting for the pipeline requirement buried in a reference file.

Sixty Thousand Words of Instructions

The depth surprised me, so I measured it.

The numbers

The 31 skills run 6,842 lines totalling roughly 59,870 words. Another 89 reference files sit in subfolders, loading only when a task calls for them. The render pipeline migration skill alone runs 543 lines with six references attached. LevelPlay carries twelve.

Most community skill collections ship a few hundred lines total. This is a technical manual rebuilt for machine consumption.

How they’re constructed

The craft shows. Skills open by detecting which of Unity’s competing implementations your project uses, because guessing wrong there produces confidently broken code. Several run a diagnostic checklist before proposing any change at all. Two include an explicit section on when the agent should refuse to use that skill, a discipline almost nobody writing skills bothers with.

The pixel perfect skill opens by warning that two entirely separate implementations exist, then branches on your render pipeline. That one decision point probably kills the most common wrong answer a model gives about blurry sprites.

So the efficiency claim starts looking plausible without any benchmark. Not because Unity asserts it. Because the failure modes these files head off are the ones that eat an afternoon.

The Skill List Is a Map of Where People Get Stuck

Line the 31 skills up by category then a picture forms that Unity probably didn’t intend to publish.

Seven skills cover 2D work plus tilemaps. That’s the single biggest cluster, larger than user interface, larger than rendering. Three separate skills handle rule tiles alone. Somebody at Unity has watched an enormous number of developers fail at the same tilemap task, repeatedly, for years.

Four skills cover user interface, which sounds reasonable until you notice why. Unity ships three competing UI systems. The skills exist partly to stop an agent picking the wrong one, a problem Unity created for itself then documented its way around.

Monetization got the deepest documentation

Then there’s the monetization cluster. In-app purchases, LevelPlay integration plus a live operations skill for running a game after launch. Three skills out of 31 pointed squarely at revenue, holding the deepest reference folders in the entire collection. LevelPlay carries twelve reference files. No rendering skill comes close.

Unity Ads sits inside LevelPlay. Unity IAP takes a cut through Unity’s own commerce stack. So the three skills carrying the heaviest documentation are also the three tied most directly to Unity’s revenue rather than yours. That may simply reflect where support demand concentrates. It also happens to be where a well informed agent earns Unity money.

Read that distribution as a support ticket histogram and it becomes the most honest document Unity published today. These are the places where developers stall, where the docs fail, then where Unity loses money when somebody gives up. Nobody writes 543 lines about render pipeline migration for fun.

It also explains the timing. Agent skills are cheaper than fixing the underlying complexity. You cannot easily collapse three UI systems into one after a decade of shipped projects. You can write four skills teaching a model which one your project already uses.

Unity Didn’t Pick Claude

The announcement positions Claude Code as the first agent to receive the plugin, with others arriving later this year.

Meanwhile the README already carries install commands for Codex plus Grok.

The version numbers give it away

Dig one level further and it gets harder to defend. The repo carries separate manifests for each agent. Claude Code’s manifest reads 0.1.2-beta.

Codex’s manifest reads 0.1.3-beta.

Codex is a version ahead. Whatever “later this year” means, it doesn’t mean unstarted, unwritten, or even untested. Unity built this cross agent from the beginning, then framed launch day around whichever directory would carry the announcement furthest.

What the packaging tells you

The repo holds three parallel plugin folders. One for Claude, one for Codex, then a generic agents folder holding its own marketplace manifest. Same skills underneath, three wrappers on top.

That structure is the interesting part for anyone outside game development. Skills started as a Claude convention. They’re becoming a portable format, with vendors shipping one body of instructions wrapped for whichever agent a developer happens to run. Grok even gets a note explaining that it demands explicit trust before installing from a repository.

Read today’s coverage as a signal about Anthropic’s position in game development and you’ll read it wrong. Unity refused to pick a side, which is correct for a company whose customers already split across four agents. The directory listing is distribution, not partnership.

About That Security Review

Unity lists a security review among its reasons for building this in house. Community skills get no such review, which is a fair point on its face.

What the files declare

Agent skills can narrow what an agent is permitted to do while running them, through a tools declaration in the file header. Three of the 31 skills use it. The unity-cli skill declares Bash, which makes sense given it drives a command line tool.

The other 28 declare nothing at all.

What that means in practice

Declaring nothing isn’t a vulnerability. It means those skills inherit whatever permissions your agent already runs with, so the setting lands wherever you left it. Most community skills behave identically, so Unity is not worse here.

But a security review is the specific claim Unity made. Scoping tool access is the most obvious control this format offers. Twenty eight skills skip it. If the review happened, it evidently didn’t cover permissions, which leaves the claim resting entirely on content rather than execution.

Ask Unity what the review covered before repeating the line. Nobody has published that answer yet.

Why Unity Stopped Leaving This to Volunteers

Unity’s stated reasoning runs more candid than most launch posts, which deserves credit.

The rot problem

Unity says plainly that community plugins exist, that they offer plenty of helpful skills, then that many are widely used. No dismissal anywhere. What Unity claims to add is authorship by the teams owning each feature, a security review, plus documentation updating whenever the engine does.

That last item carries the weight. Community plugins rot. Somebody writes a brilliant collection against Unity 6.0, gets busy, then stops updating when 6.2 changes an API. Your agent keeps producing code against a signature that stopped existing months ago, confidently, every time.

The efficiency case

Unity’s practical position is that a skill guided agent makes fewer errors, burns fewer tokens then finishes in fewer turns than one improvising through an unfamiliar API. No benchmark accompanies the claim, so it stays filed under vendor reported.

The logic tracks with what anyone running a serious Claude Code setup already knows. Skills are cheap context. They stop a model inventing method signatures, which is the most expensive failure mode in agent assisted coding because you only find out at compile time.

What’s new is a platform vendor treating that upkeep as its own responsibility. Until now the industry assumed the ecosystem would sort it out. Unity decided otherwise, which puts quiet pressure on Unreal plus Godot to answer. We covered the moment skills got a distribution layer back when the open question was whether anyone would publish to it. A company with roughly 60% of the mobile game market just answered.

What You Should Actually Do

Installing it in about a minute

Open a Claude Code session in your Unity project. Run the marketplace add command pointing at Unity-Technologies/unity-agent-plugin, then the install command for unity@unity-agent-plugin. Both are slash commands, so they belong inside the session rather than your terminal.

Prefer the terminal? The claude CLI takes the same two commands. Installs done that way load next time Claude Code starts, or immediately after you run the reload command in an open session.

Verify by typing /unity: then watching the skills appear. The plugin menu should show unity as installed while enabled. If the marketplace route fails, clone the repo then symlink it into your personal skills directory instead, which loads it in every project from your next session onward.

Who should bother

Ship in Unity 6 while already using Claude Code? Install it today. The skills load automatically once Claude spots a Unity project, they cost nothing, then worst case they sit unused. A bad skill wastes one prompt. That’s the whole downside.

Sitting on Unity 2022 or older? Skip it until you upgrade. Nothing here works for you.

Want live Editor control? Budget twenty minutes rather than two commands. You need the Unity CLI plus the com.unity.pipeline package before the MCP step does anything.

Hoping to study how a professional team writes skills then apply that to your own domain? Read them freely. Just understand the license means you’re reading for education rather than reuse.

Newer to all of this? Note that the plugin assumes you already run Claude Code with a paid plan behind it. Our Claude Pro review covers whether that subscription earns its keep before you start stacking game engines on top. For the wider map of how agent tooling fits together, the OpenClaw deep dive handles the rest.

One more thing. That version number says beta for a reason. Keep scene files in source control. Don’t let it near a prefab until you’ve watched it work on something you can afford to lose.

The Part Worth Keeping

Install it anyway.

That’s the odd conclusion, but it holds. The skills are careful. They were written by people who clearly answered the support tickets, they head off the exact mistakes that cost you an afternoon, then they take sixty seconds to add.

None of that changes what sits in the in-app purchases folder.

Unity built something useful, wrapped a restricted license around it without saying so, then tucked a competitor removal guide inside the part that touches money. All three decisions are legal. All three went unmentioned. The agent repeating them back to you sounds exactly like the one you already trust, because it is.

So install it. Read the reference folder first.

Charts and Blocks

Unity plugin skills by category

Official plugin vs community plugin