How to use Claude Code with Roblox Studio

What you need first

• Claude Code, signed in with your own Claude account. Baseplate drives the CLI you already pay for; it does not resell tokens or meter prompts.
• Roblox Studio, recent enough to have Assistant → Manage MCP Servers.
• Rojo, if you want file edits to land in the open place automatically: rokit add rojo-rbx/rojo.
• A Rojo-style project folder — a default.project.json and a src/ tree. Baseplate can scaffold one if you are starting fresh.

Step 1 — install the bridge plugin

The plugin is the half of the bridge that lives inside Studio. It long-polls the local Baseplate server, which means Studio never opens a port of its own — the same pattern Roblox uses in its official Studio MCP server. Baseplate installs it into your local plugins folder for you and tells you when a newer version ships.

Step 2 — restart Studio completely and approve HTTP

This is the step almost everyone loses twenty minutes to. A freshly installed plugin shows as installed but stays disconnected until Studio is fully restarted — closing the place is not enough, quit the whole application — and until you approve the plugin's HTTP permission prompt. If your checklist says “installed but not connected”, this is the fix, not a broken install.

Step 3 — enable Studio as an MCP server

This is the second step that traps people, because nothing about it is discoverable. In Studio, open Assistant → Manage MCP Servers and turn on Enable Studio as MCP server, then restart Studio again. This is what gives an agent Roblox's own tools — running Luau in the live session, capturing the screen, starting and stopping a playtest.

Step 4 — point Claude Code at the bridge over MCP

Claude Code reads a .mcp.json in the workspace. Baseplate writes that file and launches the CLI with --mcp-config, so the tools are present in the very first turn rather than after a reconnect. Doing it by hand looks like this:

{
  "mcpServers": {
    "baseplate": {
      "type": "http",
      "url": "http://127.0.0.1:34901/mcp?ws=<url-encoded-workspace-path>"
    }
  }
}

Codex and Cursor keep MCP configuration somewhere else — Codex in a global ~/.codex/config.toml written by codex mcp add, Cursor in .cursor/mcp.json. All three end up with the same tools.

Step 5 — start Rojo

With rojo serve running and the place connected, an edit Claude Code makes in src/ is in the running game moments later. That is what turns “here is a patch” into “here is a fix I just watched work”.

What actually changes once it is connected

The interesting part is not that the agent writes Luau faster. It is that the agent can now check itself:

1. It runs the game and watches the console — the same Output you would read.
2. It reads the real stack trace and finds the cause in your source, instead of guessing from the symptom.
3. It edits src/, and Rojo syncs the change into the open session.
4. It playtests again, and only reports PASS when the console is genuinely clean.

That loop is the whole point. A terminal agent that cannot run your game will confidently tell you a fix “should work now”; an agent that can press Play has to prove it. If you want the failure modes that survive code review but not a real Play session, the runtime guide walks through four of them with copyable tests.

Honest limits

• You still need your own Claude subscription. Baseplate is orchestration, not an AI plan, and Claude Code bills through your account.
• Verified playtests need a real Studio session open. There is no headless Roblox; if Studio is closed, you get static analysis, not evidence.
• Not every CLI can prove a PASS. Claude Code, Codex and Cursor emit checkable tool-call events. Grok Build runs fine as a builder tile but does not expose them, and Gemini CLI has no verified structured mode.
• Review what it writes before you ship. An agent that fixed a runtime error has not audited your security; the RemoteEvent checklist is the pass to run before release.