What I’d Actually Let an AI Agent Do in My Roblox Game
Back in the summer of 2024 I wrote a couple of Roblox posts about GUI stuff. How to scale a Frame so it doesn’t fall apart on a phone. How to hide the built-in CoreGui so my little Webtoon Reader place looked clean. Very unglamorous. AutoScale Lite, AnchorPoint at (0.5, 0.5), that kind of thing.
What I remember most from that era isn’t the scripting though. It’s the tab switching. Build in Studio. Alt-tab to the Creator Dashboard to make a game pass. Upload an icon. Go back. Copy the ID. Paste it into a script. Publish. Go check if anyone actually bought it. Half of “making a game” was being my own back office.
So when Roblox posted on November 4 that the Studio MCP server can now call Open Cloud, I read it twice. Then I read the docs page, which honestly is the more interesting document.
What actually shipped
If you haven’t been following, Studio has had an MCP server for a while. That’s what lets an outside agent like Claude Code, Codex or Cursor talk to the place you have open. Write a script, insert a part, run a playtest.
The DevForum announcement pushes that agent out of the editor and into the business side of your game. There’s a new tool called open_cloud_invoke, and from the same chat where the agent is editing your place, it can now:
- Create developer products and game passes, with prices and icons, from one prompt
- Change game and place settings, publish, and restart servers
- Read your analytics after an update goes out
- Import a model from Blender or Maya, then fix the scale, the missing textures and the anchoring
The Rolimon’s account summed it up for the trading crowd the same day. Over in the DevForum thread, the replies split about how you’d guess. People asked for multiple agent viewports and complained about subagents fighting over the MCP connection. A few others just asked how to hide the new MCP status indicator in the corner of Studio, because they don’t use AI and don’t want to look at it. Fair.
My favorite take was this one from @shadowaguy: an agent with permission to create game passes “will have twelve tiers of VIP carpet live before lunch.”
That’s a joke. It’s also kind of the whole article.
The permissions tab is the feature
Here’s the part I think matters more than the announcement headline. You don’t hand the agent your Open Cloud API key. Studio holds the credentials, the agent never sees them, and you decide per game what it’s allowed to touch.
To find it, click MCP in the bottom-right corner of Studio and pick the MCP Permissions tab. Or open Assistant, hit the three dots, Settings, MCP Permissions. Permissions save to your account for each game, and collaborators set their own. Me turning something on doesn’t turn it on for my friend who also edits the place.
There are four starting levels: Safe, Balanced, Full access, and Custom, which is what it shows once you change anything by hand. According to the docs, every game starts at Balanced. That means low and moderate risk scopes are on by default. Full access is the one that includes scopes that can cost Robux, affect your players, or publish to your live game.
The sections are grouped in plain English, which I appreciate. Look at your game. Edit your game. Players & social. Save & load data. Connect to the internet. Publish & manage assets. Money & monetization. Player consequences, which is bans. Security & credentials.
And the docs are pretty blunt about which ones are dangerous. Money & monetization, Player consequences, Security & credentials, plus the scopes that write game settings and publish places. Roblox’s own advice is to turn those on only for the task that needs them and turn them off when you’re done.
@OfficialZidiam put it well: this doesn’t give an agent full access to your game, it’s “only as risky as the access you give it.” Which is true. It also means the risk moved from Roblox’s side to the screen where you click toggles at 1am because the agent asked nicely.
What I’d turn on, and what I wouldn’t
I haven’t wired this into a live game yet, so take this as the plan I’d follow, not a war story. But I’ve spent enough time watching coding agents do confident, wrong things that I have opinions.
On from day one: Look at your game. Read settings, version history, analytics. This is the boring superpower. “I pushed an update Tuesday, did session length go up or down?” is a question I’d love to ask in the same chat where I made the change. Gabriel Espinosa replied with basically the same rule I’d use: start agents on read-only analytics and only give write access once you trust the setup.
On, with a folder: Publish & manage assets. The Blender import flow is the thing I’d actually use the most. Anyone who’s imported an FBX into Studio knows the ritual. It’s huge, or it’s tiny, or it’s rotated 90 degrees, or it’s suddenly grey because a texture didn’t come along. Having the agent compare against the source and fix that is real time saved. Two catches from the docs though. You approve which folders it can upload from, and it can’t sneak files into an approved folder to get around that. And uploads are permanent assets owned by the game owner. Open Cloud can’t archive or delete them, they go through moderation like anything else, and you’re responsible for what gets uploaded on your account, including what the agent uploads for you. So keep that export folder clean.
On only in a test game: Save & load data. This one surprised me. The docs say Studio uses your live data stores, and writes from your client change real player data. Their suggestion is to test against a separate game first. I’d go further. I’d make a copy of the place as its own game just for agent work and never enable data writes on the real one. If an agent “cleans up” a player’s inventory key in a test place, that’s a funny screenshot. In your actual game, it’s a refund ticket and a kid who lost their pets.
Off unless I’m sitting there: Money & monetization. I don’t think the agent would do something evil. I think it’d do something eager. You ask for “a starter pack and a VIP pass” and it helpfully makes five, prices them by vibes, and generates icons that look like a 2009 forum signature. Roblox’s own Skills tell agents to ask before using permissions that involve money, players, secrets, or publishing, and not to work around a permission you turned off. That’s good. I still want the toggle off when I’m not watching.
Get Nithilan Kumaran’s stories in your inbox
Join Medium for free to get updates from this writer.
Off, full stop: Player consequences. This is the scope that can create and remove bans. I can’t think of a solo dev workflow where I want a language model deciding who’s banned from my game. Maybe a big studio with a real moderation pipeline has a use for it. I don’t.
Edit your game and publishing: on when I’m shipping, off after. Restarting servers and publishing from a chat is convenient right up until it isn’t. If you’ve ever published the wrong place version on a Friday night, you know. I’d treat this like a deploy key. Turn it on, ship, turn it off.
One small known issue from the Nov 5 update, in case you hit it: in a place you haven’t saved to Roblox yet, execute_luau can't make HTTP requests, and you can't add domains to the allow list until the place is saved. The fix for now is File, Save to Roblox, then add your domains.
Why Roblox is doing this now
This didn’t come out of nowhere. Roblox’s Q3 shareholder letter came out a week earlier, on October 29, and it’s basically an AI letter with some financials attached.
Some numbers from it:
- Over the past month, more than 80% of the top 100 games by hours engaged used AI tools to accelerate creation
- Daily asset uploads to the Roblox cloud tripled month over month
- Since Q2, time spent in AI tools tripled among top creators, and the number using the MCP server daily grew nearly 5x
- Break and Steal an Egg and Huss Valley were both made almost entirely with AI and both hit the top 20 by concurrent players
- Build, the prompt-to-game tool, has been in alpha in New Zealand, Singapore, Serbia and Thailand. More than 500,000 people used it and published over 70,000 games
KreekCraft’s post about the letter is the one most people saw, and the replies were a lot of “Roblox devs are cooked.” Small thing, but his post says “Steal an Egg,” and the letter actually says “Break and Steal an Egg.” Those are different games. Worth getting right before we decide which one AI made.
The more useful reply I found was from @Moriarty_tech, pointing out that the 80% stat is “used AI tools to accelerate creation,” which “can be one script helper.” That’s how I read it too. If you used Assistant to write one remote event handler last month, you’re probably in that 80%. It’s a real number. It’s just not “80% of top games were made by AI.”
Jon Jordan at Regen Raygun pulled out the stat I think is the actually big one. More than 75% of Build users had never opened Roblox Studio. That’s not existing devs getting faster. That’s new people showing up who never would have made a game. He also pointed out that the same letter has daily users down 9% year over year to 138 million and hours down 15%, so 70,000 new games are landing in a smaller pool of attention. And Build will have a free base tier with paid tiers for more prompts, which means making things on Roblox starts to have a cost for the first time. On the earnings call, David Baszucki said a broader US release of Build lands in December, with a global rollout in 2027.
Put those together and the November 4 update makes a lot of sense. Build is for people who’ll never open Studio. MCP plus Open Cloud is for people who live in Studio and are tired of the Creator Hub tab. Roblox wants AI on both ends.
So what do I actually think
I’m into it. I mean that. The worst part of solo Roblox dev was never the Luau. It was the admin. Making passes, uploading icons, fixing imports, checking if an update hurt retention. If an agent can do the clerical stuff while I work on whether the game is fun, I’ll take it.
But I wrote a whole post last month about harnesses, the setup around a coding agent, and this is the same lesson in a Roblox costume. The model isn’t the scary part. The scary part is what it’s allowed to do when it’s wrong. Roblox actually did the responsible thing here. Per game permissions, no raw keys, skills that tell the agent to ask first, warnings that say “live data” in plain words. The part they can’t do for you is resist clicking Full access because the agent keeps asking for one more scope.
So my setup would be simple. A test copy of the game for anything that writes data. Read access on the real one. Asset uploads from one export folder. Money, bans and publishing off unless I’m in the chair. And if I ever look up and see twelve tiers of VIP carpet in my store, that’s on me, not the agent.
If you’ve already hooked Claude Code or Cursor up to Studio, I’d genuinely like to know what you turned on. And what you turned back off.
If You Wish To Support Me As A Creator
- Clap 20 times for this story
- Leave a comment telling me your thoughts
Thank you! These tiny actions go a long way, and I really appreciate it!
Website: https://nithilan.dev/
LinkedIn: https://www.linkedin.com/in/nithilan-kumaran/
Github: https://github.com/sathariels
