Studio MCP Server Updates and External LLM Support for Assistant

[avienn_a]

Key Takeaways

We’ve added new tools to the Studio MCP Server that will allow agents to iteratively plan, write, test, and modify your game. You can also connect Assistant to your preferred LLM by providing an API key.

---

[Update] February 25, 2026

@avienn_a

Thank you for trying out these tools and sharing your feedback.

We’ve updated the Studio MCP server to address issues where agents could fall out of sync or get stuck in loops. To ensure more reliable state management, tool descriptions now explicitly instruct agents to verify the current state before switching to Play mode and to perform a hard reset if the state is unexpected.

Key Technical Fixes in v0.2.365

• Fixed the auto-stop failure and issues where agents repeatedly entered run_script_in_play_mode.
• Since run_script_in_play_mode is only callable in Edit mode, it now always triggers a stop/reset internally to ensure a clean run.

---

We are excited to announce some significant updates that will accelerate creators who are interested in using AI as part of their development process.

Enhanced MCP Server: Better Tools for AI-Driven Iteration

Our open source Studio MCP server provides the foundation for accessing and editing your experience from external AI coding tools like Cursor and Claude Code. Today we’re introducing a suite of new tools in our MCP server designed to improve the iteration cycle.

This update includes the following new tools:

• get_console_output: Reads the output logs within a playtest session
• start_stop_play: Manages starting and stopping play mode in Roblox Studio
• run_script_in_play_mode: Runs a script in play mode and automatically stops play after the script finishes or times out. Returns structured output including logs, errors, and duration.
• get_studio_mode: Gets the current Studio mode (start_play, run_server, or stop).

When combined with the existing MCP tools, AI agents can iteratively plan, write, test, and modify code and data with less intervention and effort from creators. This capability is available today through MCP and will be fully supported in Studio Assistant in early March. In the future, we plan to integrate the MCP Server into Studio to streamline access and allow us to add new capabilities to MCP clients and Studio in lockstep.

Here’s a sped up video of using Claude Desktop with the MCP Server to automatically test and fix a feature.

Assistant Flexibility: Integrate External LLMs

We are also increasing the flexibility and power of Assistant.

As a first step, we’ve added the ability to connect and utilize external Large Language Models (LLMs) to power Assistant’s generative capabilities.

You can now leverage models from the following 3rd-parties:

• Anthropic
• OpenAI
• Google Gemini

Follow the links above to learn how to get an API key for each platform and see the setup instructions at the end of this post to get started.

Next up, we’ll be adding support for connecting Assistant to third-party MCP servers, enabling access to tools like Blender and Figma. This will allow you to create a UI in Figma, and then ask Assistant to recreate that UI, generating elements directly in your experience.

What’s next

Throughout the coming year, we will be rolling out a series of enhancements to both Assistant and our MCP server. Our goal is simple: to provide you with a more integrated, creative, and productive experience every time you open Studio.

We welcome your contributions and want your feedback!

Please let us know If you run into any issues using the MCP server or setting up your own API Keys. We’d also like to hear about additional API Key providers and third-party MCP servers you would find most useful.

Happy creating!
The Roblox Studio Team

How to add an API key

Click here for details

1. Open the Menu: Click the three dots… in the top-right corner → Manage API Keys

   

2. Select Provider: Choose a provider, then paste your API key

   • Security: Your key is masked in the UI and stored securely on your device
   

3. Choose your model: Enable the provider and select a model from the Model dropdown to start using it right away

   

[Rixxles]

AI is supposed to be a tool, you’re trying to replace error finding and scripting entirely which means beginner scripters using it will never learn or want to do it themselves.

Please just stop incorporating AI into everything to make your investors happy, focus on the players instead..

[ShoomamaAtlantis (Shoomama)]

Great work from the AI team! Being able to run play mode via MCP is massive for quality assurance and helping ensure updates have minimal bugs on launch. Looking forward to testing this stuff out.

Do you guys think you’d be able to add support for Claude Code and/or Codex subscriptions? API pricing for AI is fairly expensive at this moment. Also, support for “one API key” providers like OpenRouter would be nice to have.

[Mythiclys]

Anything related to A.I. and basically stopping people from learning how to code (Or learning anything) themselves, Gross. Get it out of here. We don’t need A.I. slop in EVERYTHING, Even better, just remove it from the platform!! Yeaahhh… Oh wait, you can’t because of Investors. Yuck.

[Perthyz]

I think this will be useful for rapid iterative development, rather than a complete replacement for programmers overall. Definitely something to keep track of

[Clxzed]

Thanks for making updates that substitute knowledge for ignorance! I’m so glad to live in an era where I don’t have to learn anything, and I can trust my virtual word predictor to do the learning for me.

[TheNexusAvenger (Nexus)]

@avienn_a

As a first step, we’ve added the ability to connect and utilize external Large Language Models (LLMs) to power Assistant’s generative capabilities.

You can now leverage models from the following 3rd-parties:

• Anthropic
• OpenAI
• Google Gemini

Are there any plans to allow for custom LLM hostings that are API-compatible with the supported systems? For example, self-hosted models through Ollama or LM Studio. There are a bunch of interesting developments with open-weight models, but the model host limitations prevent them from being experimented with or used.

[megasuperalfonato (RonaldoR9)]

I feel roblox is getting in love with AI too much

instead of this I think Roblox should work on diverse feature requests like better graphic options or even really minor ones like water on baseparts or also revamping terrain and terrain colors per zone, etc! There’s many many more

[N_aps]

sick been advertising this a ton on X, will tell more people about it keep it up

[DivineChad (BOBOH)]

The use case of AI here is to be a tool,
this just makes the workflow easier for debugging with the help of LLMs.
I don’t think most beginners have access to frontier models the bridge is supposed to be used with.
The MCP Tool is very useful, believe it or not majority of developers use AI in some kind of form already, vibe coding and wasting budget on it isn’t what this is meant for.

[GoobsterX (Diamond)]

While I believe that A.I. should not be a replacement for real developers coding experiences, this could definitely be useful tool to assist in the smaller parts of the development process. For the times where you just need to quickly debug something minor, get insight, or help inexperienced developers learn, this is definitely an interesting way to do it! I’ll personally be testing it out myself for when I already rarely use AI, it’s a nice and quick integration, but please don’t go around forcing AI everywhere or doing anything more than this.
This can be nice feature to have - but it should not go too far where it is trying to replace anything, and developers should still be doing development using their own skills, to learn the systems and make an authentic game.

[Rixxles]

This is literally just vibe coding. You’re too lazy to find and fix an error yourself, so you just use the AI to do it for you. That’s not you doing it, and you’ll never learn how to find bugs.

If you do it once and you realise how much faster it is, you’ll rely on it even for small bugs, and the ONLY use for AI should be if you can’t find the bug, no matter how much you look.

[DivineChad (BOBOH)]

Sometimes you have more important things to do than to search a billion years for a bug where you certainly know the output isn’t right? That doesn’t mean I can’t find it myself, these AI tools are mainly used for fast prototyping and writing unit tests if your’e brave… it has nothing to do with taking shortcuts as a beginner… a beginner shouldn’t be using these tools in the first place, I doubt they can even get a working development env, setup.

[PeZsmistic]

Interested in support for locally run LLMs via Ollama, are you guys looking into this, or just focusing on cloud LLM providers for now?

[Bloxec (一一一一)]

Glad to see you guys working on improving the built-in Assistant.

Since you’re already working on supporting external providers, I’d really like support for a way to plug in self‑hosted models (Ollama/vLLM) via a generic HTTP or MCP bridge so folks using self‑hosted setups can still use Assistant with their own models.

On top of that, one cloud provider I’d love to see natively supported is OpenRouter, since it gives access to many different models behind a single API and billing setup. In my experience, that kind of flexibility lets you quickly swap between fast/cheaper models for quick edits vs. frontier models for big refactors.

[Rixxles]

If you don’t have the knowledge or don’t want to spend time looking for a bug in your OWN code then either you didn’t code it, or you aren’t good enough at scripting.

[arox1007 (arox)]

This is amazing! e.e already installing it

[A_Mp5 (amp)]

this is not making investors happy, this is making cracked developers happy. like me

[DivineChad (BOBOH)]

If youre on a tight schedule and already need to do a bunch of code reviews, a little bit of help from AI for rudementary task doesn’t hurt. Using AI in any form for development is no indicator for a skill level in programming. Your’e sight on AI is too boxed and naive.