Free & open source · MIT

Let your AI agent build inside Roblox Studio

An open-source Model Context Protocol server that gives Claude, Codex, Cursor & Antigravity hands-on control of a live Roblox Studio session. It runs Luau, builds UI, terrain and lighting, screenshots the viewport, and keeps a two-way Studio ↔ disk mirror.

npm version npm downloads per month MIT license Model Context Protocol server
# register with your AI agent; it launches the server for you
claude mcp add roblox-mcp-pro -- npx -y roblox-mcp-pro@latest

Just talk to your agent

Connect the server, click the MCP button in Studio, then describe what you want. The agent picks the right tools, runs them in your open session, screenshots the result to check its own work, and iterates while you watch.

Build a glassy settings panel with a volume slider and a close button, then animate it sliding in.

Scatter 30 trees across the terrain, avoiding the water, and add soft evening lighting with fog.

Read this screenshot and rebuild this shop UI pixel-for-pixel.

Sync the whole place to disk so I can edit scripts in my editor, two-way and live.

Built for real Studio work

29 focused tools, not one giant "run code" hammer.

29 focused tools

Instances, properties, scripts, raw Luau, UI, terrain, lighting, effects, tweens, camera, animation, audio, physics, marketplace assets, spatial queries, bulk edits, playtests, screenshots.

Two-way sync

Mirror the live DataModel to local files and back. Edit scripts in your own editor; changes flow both ways. Multiple Places sync at once, each on its own engine.

Multi-agent

Claude Code, Codex and Antigravity can all drive the same Studio session concurrently through one shared broker.

It sees what it builds

capture_studio screenshots the real viewport, so the agent closes the loop instead of building blind.

Live dashboard

Watch connected agents, plugin status, the command queue and sync state in real time at 127.0.0.1:3690.

Zero maintenance

npx … @latest auto-updates the server; the Studio plugin and agent skills self-install on startup.

Up and running in two steps

You never start the server by hand; your AI agent launches it with the command you register.

claude mcp add roblox-mcp-pro -- npx -y roblox-mcp-pro@latest
1

Register the server

Use the command above for your client. npx downloads & runs the package on demand (no clone, no build). @latest keeps it auto-updated.

2

Install the Studio plugin

Auto-installs on first run, or do it now: npx roblox-mcp-pro@latest install-plugin. Windows → %LOCALAPPDATA%\Roblox\Plugins, macOS → ~/Documents/Roblox/Plugins.

3

Connect & verify

In Studio click the Roblox MCP Pro toolbar button (so it's highlighted), then ask your agent to call system_info. You should see pluginConnected: true.

Optional environment variables

VariableDefaultMeaning
ROBLOX_MCP_PORT3690Bridge port (the plugin must match).
ROBLOX_MCP_TOKENnoneShared secret; also set it in the plugin.
ROBLOX_MCP_NO_PLUGIN_AUTOINSTALLoff1 disables auto-installing the plugin.
ROBLOX_MCP_NO_SKILL_AUTOINSTALLoff1 disables auto-installing agent skills.
ROBLOX_MCP_NO_OPEN_DASHBOARDoff1 stops the dashboard opening on startup.

29 tools

Each does one job well; the agent composes them.

Core

Read and change the DataModel.

system_infoConnection + Studio session status.
execute_luauRun Luau in Studio; capture output & return values.
query_instancesSearch the DataModel; serialize matches.
find_instancesTargeted multi-filter search; grouped results.
scene_overviewOne-call map: class histogram + shallow tree.
describe_instanceProps + children + ancestry, or class reflection.
mutate_instancesCreate / edit / move / clone / delete instances.
manage_propertiesProperties, attributes & CollectionService tags.
batch_executeMany operations in one round-trip + one undo.

World & visuals

Shape the scene and how it looks.

manage_terrainVoxel fills, ramps, material swap, smooth, hills.
manage_lightingTime of day, fog, ambient, exposure, shadows.
manage_effectsBloom, Blur, ColorCorrection, DoF, Atmosphere, Sky.
manage_cameraRead, set, or aim the CurrentCamera.
manage_tweenAnimate properties with TweenService.
manage_physicsPhysical properties and welds.
spatial_queryBox / radius / raycast / nearest / placement checks.

Content

Author UI, audio, scripts and assets.

manage_uiBuild/edit GUI hierarchies (UDim2/Color coercion).
ui_previewShow a GUI on a clean backdrop for comparison.
manage_audioCreate, configure & preview Sounds.
manage_animationManage Animation instances; best-effort preview.
manage_assetsSearch marketplace; insert assets; product info.
manage_scriptsRead/write/create scripts; line edits, find-replace.

Sync, sessions & Studio

Move data to disk, route agents, run playtests.

manage_syncTwo-way Studio ↔ local mirror; multiple Places at once.
manage_agentsList connected agents/Places; bind to a Place.
manage_selectionRead/change Explorer selection; watch user clicks.
manage_studioStudio info + playtests (Run / Play Solo / multiplayer).
manage_logsRecent Output log history (filter by type / since).
workspace_stateSession snapshot + tree diff since last check.
capture_studioScreenshot the Studio window so the agent sees the render.

How it works

Three processes, two hops, and several agents on one Studio session.

AI Agent A ─MCP stdio─▶ MCP client ─┐
AI Agent B ─MCP stdio─▶ MCP client ─┼─HTTP 127.0.0.1:3690─▶ Broker ◀─long-poll─ Studio Plugin (Luau)
AI Agent C ─MCP stdio─▶ MCP client ─┘                       queue + dashboard    dispatch handlers
1

MCP client

Exposes the tools over stdio. Each AI agent spawns its own.

2

Broker

One shared localhost process owns port 3690, queues commands, talks to the plugin. The first client auto-spawns it; the rest just connect, so several agents drive one Studio session at once.

3

Studio plugin

Long-polls the broker, runs each command in Studio, and posts the result back.