Assistant Updates: Mesh Generation, New MCP Server Tools, Screenshot Tool, and More

Updates Announcements
, , ,
1

Hi Creators,

We’re excited to announce more updates to Mesh Generation and introduce new Screenshot and MCP tools!

More Control to Mesh Generation

This update allows you to generate a textured 3D mesh from a text prompt in seconds using the GenerateModelAsync API, in Studio. Whether you need a specific prop for a scene or a placeholder asset for a new game mechanic, Assistant can now help create and iterate faster.

How to Generate

There are two ways to start creating:

  1. Direct Commands: Use the /generate command followed by your prompt (e.g., /generate a weathered stone well).

  2. Conversational Requests: Ask Assistant directly, for example “make a futuristic crate”

Improved Workflows

We’ve improved workflow efficiency by adding:

  • Batch Creation: Users can now simultaneously generate multiple meshes, ideal for bulk creation.

  • Reusability: A single generated asset can be instantly applied to multiple locations with the click of a button.

  • Multitasking Support: Users can continue with other tasks while Assistant is generating a mesh, minimizing downtime.

Generate an oak tree, birch tree and pine tree, then make a forest out of them

Granular Control

This update introduces more precise control over the generated meshes:

  • Bounding Boxes: Use a Part in your workspace as a bounding box. Select a part before prompting and Assistant will utilize its size and location as input, ensuring the generated mesh fits within the defined space.

  • Max Triangle Count: Define a maximum triangle count for the returned model. Lower values result in more faceted and low-poly generations. This is an optional input with the default set to 10000.

    • For smaller, simpler objects, or if you are going for a low-poly aesthetic, we recommend a lower triangle count (100s to 1K).
    • For larger or more detailed objects, you can use the default or explore reducing it based on your overall scene complexity. This doc has helpful guidelines for optimizing for performance.

Assistant-mesh-gen

Generate a tent and a campfire with bounding boxes (video sped up for demo purpose)

New Tools in Assistant and MCP Server

We’re continuing to expand the capabilities available through Assistant and the Studio MCP Server. This update adds the following tools, available both in Studio and to external AI clients like Claude, Cursor, and others:

  • insert_from_creator_store: Search for and insert models from the Creator Store directly into your experience. This lets AI agents pull in community-created assets, plugins, and models without manual browsing.

  • generate_mesh: Generate a textured 3D mesh from a text prompt using AI. The same mesh generation capabilities available in Assistant are now accessible through any connected MCP client.

  • generate_material: Generate custom material variants from a text prompt.

These tools are available today through Assistant and the built-in Studio MCP Server.

Screenshot Tool

We’re adding a Screenshot tool to the MCP Server, also available for creators using their own API keys in Assistant. Combined with the playtest automation and virtual input tools from our last update, AI agents can now see what’s happening in your experience - not just read code and logs.

  • screen_capture: Captures the current Studio viewport in play mode and returns the image data. Agents can use this to verify visual changes, check scene layout, or inform their next steps.

OpenGameEval Update

Earlier this year we open sourced OpenGameEval, a benchmark for evaluating how well AI models perform on real game development tasks in Roblox. We’re now expanding it with debug focused evals.

  • 30 new evaluations built from 15 base scenarios, each with 1–3 injected bug variants.

  • Tests a model’s ability to identify and resolve issues within an existing game, a core skill for AI-assisted development workflows.

  • Available now in the OpenGameEval repository.

What’s Next

We are actively developing new tools to generate and iterate on 3D assets, including advanced procedural object generation and editing. We are also improving our prompt processing to ensure a more accurate understanding of your requirements before beginning the generation process.

Multiple chats and chat history are also coming soon, making it easier to keep workstreams organized, preserve context across tasks, and pick up where you left off.

Stay tuned for more updates. Please let us know if you run into any issues or have suggestions for new features you’d find useful.

Happy creating!

70 Likes
2

This topic was automatically opened after 10 minutes.

3

I think this is another solid addition to the MCP server. Asset creation has definitely been a bottleneck, so this helps unblock iteration quite a bit.

That said, I still see generated assets as more of a template/temporary solution, useful for rapid testing and prototyping, but ideally something artists should replace with a proper quality pass later.

Overall though, this is a big win.

20 Likes
5

Mesh and text generation are fun to experiment with. I’m interested in seeing how much more advanced this can get as long as they aren’t used for laziness since I think it could have some creative usage if incorporated into gameplay

3 Likes
6

This is pretty exciting stuff!

I think players will always prefer hand-made assets, purely from an artistic soul PoV, but I can see mesh gen (especially via bounding box) being extremely useful for level prototyping/getting beta versions up.

I’m sure this was the line of thinking you had when developing this, but the bounding box stuff logically follows from level greyboxing, and reduces the friction that I as a programming-first dev experience when designing levels.

I look forward to using this, it will definitely make level design easier :smiley:

6 Likes
8

I’m trying the new screen_capture tool with claude desktop opus 4.6 and I’m getting warnings for the screenshot being too large, it states the max size is 1mb.
Maybe the screenshots need more compression?

4 Likes
9

Claude has 1MB MCP tool result limitation.
Rolling out jpeg encoding, screenshot size should be about 1/3 of the png size, which make it less than 1MB in most cases.

4 Likes
10

Nice. Claude actually automatically lowered the in-game graphics setting for me which did actually reduce the image file size below 1mb

2 Likes
11

Interesting results in the posted LLM Leaderboards. I’m looking forward to running my own tests after my project concludes.

5 Likes
12

This is awesome Roblox! I’ve been waiting for improvements to MCP and mesh generation. Thank you!

3 Likes
13

This is an amazing addition to the MCP server. I originally had my own separate MCP server to feed screenshots to my AI agents, but without a native Studio implementation to capture the viewport content, it was more tedious to use. Thank you for adding this. Please keep up these useful MCP server feature additions!

4 Likes
14

Being able to hook external tooling into the assistant workflow is gonna open up alot of possibilities for automated testing and asset pipelines.

2 Likes
15

This AI slop would cost Roblox a lot of money without even generating more revenue.

3 Likes
16

This does work surprisingly well. It would be great if I could figure out how to prevent going through the MCP costing literally 4x more than just working directly through the filesystem via script sync though. Have to do this hybrid or it’s way too expensive to be practical.

3 Likes
17

Hi @PeZsmistic! Could you explain your setup a bit more? It sounds like you are describing an external model failing to edit your synced scripts directly, and going via the MCP tools (ie multi_edit) instead? Is that correct?

2 Likes
18

Even though this sucks, there is little they can do. AI right now is big in the investor market, and stopping it could lead to another big stock drop.

4 Likes
19

Coolio! I feel like this would be most useful to give players a tiny bit of creative freedom in a sandbox game or other type of game, without them having to actually make every indivudal part themselves. Although I prefer the latter, newer generations of players seem to pick the former.

I am mixed on if I’ll use this in the future, but I feel like many different games could be based around AI roleplaying and stuff like this.

1 Like
20

image
image412×272 16.7 KB

i cant create 3d meshes using assistant.
If this works i think it can be great to prototype and test ideas before investing on 3d art

2 Likes
21

I think you need to say /generate

2 Likes
22

I’ve been using MCP in Studio since day one and I’m glad to see them getting more useful over time. The team are doing a great job with this.

I did have one suggestion; I want to add a script to my game that provides instructions for the AI - similar to a Claude.md file. Whenever I get an AI to document my game it will always come up with issues that are intended behaviour. I created a module that tells the AI how to interpret the context of my game rather than having to explain it every time, but the AI usually doesn’t know that such a script exits.

Therefore my suggestion is to add instructions to the MCP that let developers define a “AI_Context” module or something that AI’s know how to access/read before they go off an explore the structure of the game. It would also improve token usage of these LLMs as they wouldn’t need to re-read entire games just to “understand” the game they are looking at.

An alternative would be some kind of tag similar to TODO that the MCP gives the LLM instructions such as

-- @AIContext This function works as intended and should not be modified
function helloWorld()
  print("goodbye!")
end

You can sort of do this already but it could be useful to formalise it with the MCP.

4 Likes