Hi Creators,

We’re back with another round of updates designed to make AI-assisted development in Studio more powerful and easier to use. Here’s what’s new.

The Data Model Search Subagent

When agents are working in a large place, they often need to run many exploration steps in a row to answer a single question about your codebase, searching the game tree, reading scripts, searching (e.g. grep) for references. That’s slow, and it fills the main conversation with low-level tool results that aren’t useful to read.

One of the most consistent pieces of feedback we’ve heard is that agent performance can drop off on longer tasks, particularly in larger places. A lot of that comes back to context management: the more intermediate tool calls pile up in the main conversation, the less room the agent has to reason about your actual question. The Data Model Search subagent is one of the ways we’re tackling this, by moving the exploration work out of the main context and into a scoped subagent, the main conversation stays focused on your task.

Today we’re introducing the Data Model Search subagent, which joins the Playtest subagent as another scoped helper. The Data Model Search subagent is accessible in Assistant and through the MCP server. When an agent needs to investigate something in your place, it can now spin up a Data Model Search subagent to handle that investigation on its own thread and return a compact summary to the main conversation.

Key properties of the Data Model Search subagent:

Runs in parallel: Multiple Data Model Search subagents can run concurrently, so agents can investigate several parts of your codebase at once instead of searching them one by one.

Clean main conversation: Only the findings come back to the main thread, so you don’t have to scroll past a wall of intermediate tool calls to see the answer.

From your side, this mostly just looks like the agent responding faster and with more specific answers when you ask questions about your place. Behind the scenes, it’s doing a lot more searching than before, just not in a way you have to watch.

Quick Connect for MCP Clients

After enabling Enable Studio as MCP server in Assistant Settings, you’ll see a new Quick connect section that lists supported clients detected on your machine. Flipping the toggle for a client handles the config insert in one-click, with no manual file editing required. Most clients would need a manual restart/mcp config refresh after that.

Supported clients for Quick connect: Antigravity, Codex CLI, Claude Code, Claude Desktop, Cursor, Gemini CLI, and Visual Studio Code.

Only clients already installed on your computer will appear in Quick connect. If you don’t see the client you want, install it first, then restart Studio so it can detect the install.

Getting set up

Install and do initial setup for the AI client you want to use (e.g. Claude Desktop)

In Assistant Settings, turn on Enable Studio as MCP server

Flip the toggle for your client under Quick connect

Restart the AI client so it picks up the new settings

Many AI clients keep running in the background after you close the main window. Check your system tray to make sure it fully restarts.

You’re connected!

For clients not listed under Quick connect, the existing Setup instructions section is still available below with the manual JSON Configuration and Startup Command, unchanged from the previous version.

What’s next

As mentioned in our last post, we’ll be rolling out regular improvements to both Assistant and our MCP Server in the coming months so stay tuned to find out more. A couple of things we’re actively working on:

Context compaction: Further improvements to how Assistant manages long-running conversations, so performance stays steady even on extended tasks. The Data Model Search subagent is one lever here, and context compaction is another

Multi-chat sessions and chat history: Run multiple Assistant sessions side by side, and browse a history of past prompts and responses so you can pick up where you left off

Copy Unique ID: Right-click any instance in the Viewport or Explorer and select Copy As → Unique ID to grab a stable, text-based handle you can paste into prompts for external LLMs via MCP. Works even when multiple siblings share the same name.

Special thanks to @swish741 @Fingolfin_123 @maggiez919 @Urukeli for their work above!

Happy creating!

The Roblox Studio Team