How to use OpenAI Codex with Roblox Studio

What you need before you start

Which Codex to use

OpenAI lists four Codex clients that sign in with ChatGPT: the ChatGPT desktop app in Codex mode, Codex CLI, the Codex IDE extension and Codex web, which hands tasks to Codex Cloud. What matters for Studio is where Codex runs: Roblox's MCP server is a local process on your computer, so Codex has to run there too.

Step 1: Install Codex CLI and sign in

Start with Codex CLI, the client on Roblox's list. These are the install commands OpenAI's CLI page gave on 11 October 2026; check OpenAI's Codex CLI page for the current ones.

curl -fsSL https://chatgpt.com/codex/install.sh | sh

powershell -ExecutionPolicy ByPass -c "irm https://chatgpt.com/codex/install.ps1 | iex"

1. Make a folder for your game's files. If you have Git, put the folder under Git: OpenAI recommends Auto mode for folders under version control and read-only for other folders, and advises Git checkpoints before and after a task so you can revert changes.
2. Run codex in that folder. The first time, choose Sign in with ChatGPT. To sign in with an API key instead, see OpenAI's authentication page.
3. Codex may start in read-only mode until you trust the folder. That is fine for now: Step 3 explains why Studio needs a setting of its own either way.

mkdir my-game
cd my-game
git init
codex

On Windows, use the native installer above rather than running Codex inside WSL for this setup. Roblox's Windows command runs cmd.exe with a %LOCALAPPDATA% path, and Roblox does not document WSL.

Step 2: Connect Codex to Studio

Turn the server on first. In Studio, open Assistant, click … > Manage MCP Servers, and turn on Enable Studio as MCP server. Keep the place you want to work on open before you start Codex: in our check with Studio closed, Roblox's StudioMCP program waited 10 seconds and then returned no tools.

Quick connect

1. In the same panel (Assistant Settings > MCP Servers), expand the Quick connect drop-down. It lists the supported clients installed on your computer.
2. Turn on Codex CLI. If it is missing, install Codex CLI and restart Studio.
3. In a terminal, run codex mcp list to see the servers Codex has configured. Roblox does not say which file quick connect writes for Codex or what it names the server, so note the name this shows.
4. Before you give Codex a task, add the approval setting from Step 3 to that entry. Open ~/.codex/config.toml, where Codex keeps MCP servers by default, find the [mcp_servers.…] header with that name, and add the single line default_tools_approval_mode = "prompt" under it, above the next line that starts with [. Do not paste a second Studio table from the manual setup below: under the same name TOML rejects it, and under another name Codex gets two Studio servers, one of them without the prompt setting.
5. Restart Codex: exit the session and run codex again. Roblox says a client needs a restart to apply changes. If you later turn quick connect off and on again, check that the line is still there.

Manual setup in config.toml

Use this instead of quick connect, not as well. Codex keeps MCP servers in ~/.codex/config.toml by default (on Windows, %USERPROFILE%\.codex\config.toml), and OpenAI says the ChatGPT desktop app, Codex CLI and the IDE extension share that configuration for the same Codex host. Each server is a [mcp_servers.<name>] table with a required command and optional args. These entries put Roblox's documented commands into that format. The name Roblox_Studio matches Roblox's own examples. Before you start Codex, run codex mcp list and check that it shows only one Studio entry.

[mcp_servers.Roblox_Studio]
command = "/Applications/RobloxStudio.app/Contents/MacOS/StudioMCP"
startup_timeout_sec = 30
default_tools_approval_mode = "prompt"

[mcp_servers.Roblox_Studio]
command = "cmd.exe"
args = ["/c", "%LOCALAPPDATA%\\Roblox\\mcp.bat"]
startup_timeout_sec = 30
default_tools_approval_mode = "prompt"

Two lines are ours, not Roblox's. startup_timeout_sec = 30 is headroom: Codex's default is 10 seconds, and we have not timed StudioMCP with Studio open, so 30 seconds leaves room over that default. It does not fix missing tools: in our check with Studio closed, StudioMCP itself stopped after 10 seconds and returned an empty tool list, which a longer wait would not change. default_tools_approval_mode = "prompt" is explained in Step 3. In the Windows entry, keep the double backslashes: in TOML a single backslash is a parse error. If the file already lists other servers, add this table after them.

Or add it with one command on macOS

On macOS, codex mcp add adds the server to your Codex configuration, which is ~/.codex/config.toml by default. The -- separates Codex's own options from the command that starts the server. Then open the file, find the new [mcp_servers.Roblox_Studio] header and add default_tools_approval_mode = "prompt" under it, plus startup_timeout_sec = 30 if you want the headroom. As with the manual entry, use this instead of quick connect, not as well.

codex mcp add Roblox_Studio -- /Applications/RobloxStudio.app/Contents/MacOS/StudioMCP

On Windows, edit the file instead. We have not checked how PowerShell passes %LOCALAPPDATA% through to that command.

Check the connection

• In a terminal, codex mcp list shows the configured servers. There should be one Studio entry.
• Inside Codex, type /mcp to see the active servers, or /mcp verbose for detailed server diagnostics.
• In Studio, Assistant > … > Manage MCP Servers shows a green indicator with the number of connected clients.
• Because the configuration is shared, OpenAI's docs mean the same server should also appear in the IDE extension and the desktop app. We have not checked this with Studio.

Step 3: Keep Codex asking before it uses Studio's tools

Set this before Codex touches your place. Codex's sandbox limits what it can do on your computer, such as which files it can write. It does not cover Studio. OpenAI's permissions page says permission profiles govern commands that run on your machine, while MCP servers use their own process or transport and their own controls. So the read-only sandbox does not stop Studio's tools. What decides whether Codex asks you first is the approval setting on the Studio server. Put it on the one Studio entry that codex mcp list shows, under that entry's name: if quick connect made the entry, add the single line default_tools_approval_mode = "prompt" to it rather than a second Studio table.

OpenAI's MCP page lists four values for default_tools_approval_mode: auto, prompt, writes and approve. It describes only writes, which prompts for tools that are not marked read-only. The entries in Step 2 use prompt so that Codex asks before Studio's tools run. Roblox does not say which of its tools are marked read-only, and we have not seen Codex's prompts with Studio, so check on your first request that Codex asks before each Studio tool call.

Optional: skip the prompt for tools that only read

Once you trust the setup, each tool can override the server's setting with its own approval_mode, and disabled_tools gives the server a deny list. Keep the server's own lines above the first tools table: in TOML, a line below a table header belongs to that table. The example shows the whole manual entry. If quick connect made yours, add only the new lines, and use the name codex mcp list shows in the tools headers too.

[mcp_servers.Roblox_Studio]
command = "/Applications/RobloxStudio.app/Contents/MacOS/StudioMCP"
startup_timeout_sec = 30
default_tools_approval_mode = "prompt"
disabled_tools = ["insert_asset", "upload_image", "generate_mesh", "generate_material", "generate_procedural_model"]

[mcp_servers.Roblox_Studio.tools.script_read]
approval_mode = "approve"

[mcp_servers.Roblox_Studio.tools.search_game_tree]
approval_mode = "approve"

The deny list names the Studio tools that bring outside content into your place, generate assets or send images to Roblox. The two approve tables cover tools that only read: a script's text and the game tree. Add other read tools the same way if you like, such as script_search, script_grep, inspect_instance and get_console_output, and leave multi_edit, execute_luau and start_stop_play on prompt. The full tool list is in the Studio MCP guide. On Windows, use the same lines with the Windows command and args.

Keep Codex's own protections too

• In Auto, Codex can read files, make edits and run commands in the working folder without asking. OpenAI recommends Auto for folders under Git and read-only for other folders.
• If you want to approve each Studio call yourself, keep automatic approval review off. OpenAI says that with it on (approvals_reviewer = "auto_review" in config.toml, or Approve for me in the desktop app or IDE extension), eligible approval requests, including side-effecting MCP tool calls, go to a reviewer agent instead of to you. The default, approvals_reviewer = "user", sends them to you.
• In the default workspace-write sandbox, .git is protected as read-only, so Codex needs your approval to commit.
• OpenAI marks --dangerously-bypass-approvals-and-sandbox (alias --yolo) as not recommended: it runs with no sandbox and no approvals. Do not use it while Studio is connected.

Step 4: Start with a request that only reads

Open the place you want to work on and close any you do not. For the first session, start Codex with the read-only sandbox and approvals on request. That limits files and commands; Studio's tools still follow the prompt setting from Step 3.

codex --sandbox read-only --ask-for-approval on-request

Then give Codex a first request that cannot change anything:

Use the Roblox Studio MCP tools. Call list_roblox_studios and tell me which place you are connected to, then list what is in Workspace. Do not change anything.

On a fresh Baseplate, Roblox's coding harness guide expects the answer to list the default Camera, Terrain, Baseplate and SpawnLocation, and notes that you might need to approve the command. Asking for list_roblox_studios first is a good habit: one client can reach every Studio window you have open, and every tool call names its target window with a studio_id.

Step 5: Give Codex your Roblox rules in AGENTS.md

Codex reads AGENTS.md files before doing any work, so your project rules go in one at the top of your project folder. Run /init inside Codex for a starter file, or write it yourself. Cursor and Claude Code can also read an AGENTS.md, so one file can serve all three.

This example was written for this guide, for a project that uses Script Sync. Change the names to match your game and cut what you do not need. The Claude Code guide explains why each Script Sync and Luau rule is there.

# Roblox project rules (example: edit to fit your game)

## Project
- Roblox game "Wave Survival". The place open in Studio is named WaveSurvival.
- Scripts sync to Studio with Script Sync:
  - ServerScriptService/ holds server game logic
  - ReplicatedStorage/ holds ModuleScripts shared by server and client
  - StarterPlayerScripts/ holds client code
- File names set the script type (Script Sync rules):
  - name.server.luau = Script with Server run context
  - name.local.luau = LocalScript
  - name.luau = ModuleScript
  - Do not use name.client.luau unless I ask. In Script Sync it makes a Script with Client run context, not a LocalScript.
- Never add attributes or tags to synced scripts. Script Sync ignores them.

## Roblox Studio MCP
- Call list_roblox_studios first. Work only in the place named WaveSurvival.
- Edit synced scripts as files, not with multi_edit.
- Use the MCP tools only for what is not a synced script: GUIs in StarterGui, RemoteEvents, parts.
- Never use insert_asset, upload_image or the generate tools unless I ask in this session.
- Ask before deleting any instance you did not create in this session.

## Luau
- Use task.wait, task.spawn and task.delay. Never wait, spawn or delay.
- Get services with game:GetService.
- The server decides coins, health, inventory and rewards. A LocalScript only asks.
- Every OnServerEvent handler checks the type and size of each argument and limits how often each player can call it.
- Wrap DataStore and HttpService calls in pcall and handle the failure.
- Never put API keys, webhook URLs or passwords in scripts or in this repository.

## Git
- Do not commit. I review the diff and commit myself.

## Before you say a task is done
- List every script and instance you changed.
- Read the Output with get_console_output and report any errors.
- Tell me what to playtest by hand.

To check that Codex picked up the file, start a session in the project folder and ask it to summarize the instructions it is following.

Step 6: Work on scripts as files (optional)

Codex works on files in a folder. Roblox's Script Sync can mirror the scripts in your place as .luau files in that folder, so Codex edits code as files, Git keeps their history and git diff shows what changed. The MCP tools are then left for what is not a script, such as GUIs, RemoteEvents and parts.

The setup steps for Script Sync and Rojo, and the file naming rules, are in how to use Cursor or VS Code with Roblox Studio. The Claude Code guide has a short Script Sync setup for an agent that works the same way for Codex.

A safe way to work: checkpoint, review, playtest

1. Checkpoint both sides. Commit your files in Git, as OpenAI advises before and after a task, and save a version of the place in Studio with File > Save to Roblox with Notes. The Studio MCP guide shows how to restore one from Version History.
2. Approve Studio tool calls one at a time, especially multi_edit, execute_luau and start_stop_play, which change your place, run code or start a playtest. That needs the prompt setting with automatic approval review off, as in Step 3.
3. Review the diff. Run git diff to see what changed in your files. For changes made in Studio, read the scripts Codex touched and check the Explorer for objects that appeared or disappeared.
4. Playtest it yourself. Through the MCP tools Codex can start a playtest and read the Output, but it only checks what it chose to check. Play the feature as a player would before you commit or publish.
5. Commit yourself once it works. In the default sandbox Codex needs your approval to write to .git, which keeps each commit in your hands.

What you can undo, and how

OpenAI's answer for file changes is Git: checkpoints before and after a task so you can revert. Changes made in Studio through MCP tools are not files in your folder, and Roblox's MCP page does not say whether they land in Studio's undo history, so for those your saved version is the way back.

Keep secrets and live data out of reach

• Do not paste API keys, Open Cloud keys, webhook URLs or passwords into a prompt or into AGENTS.md, which is usually committed to your repository.
• Work in a test copy of your game, not the live one. The Studio MCP guide explains why a playtest an agent starts can touch real player data.

What Codex costs for Roblox work

The Studio side has no documented price: the MCP server is part of Studio, and Roblox's MCP page mentions no charge. The cost is Codex's, and OpenAI offers two ways to pay (as of 11 October 2026):

• A ChatGPT plan. OpenAI's Help Center says Codex is included on every ChatGPT plan, including Free and Go. Codex Cloud is not included with Free or Go. OpenAI's pricing page describes Free and Go only in terms of the desktop app, and its feature table lists Codex CLI and the IDE extension for Plus, Pro, Business, Enterprise and Edu, with no Free or Go column. So whether a Free or Go account can use Codex CLI with Studio is not clear from OpenAI's pages, and we have not tested it.
• An OpenAI API key. OpenAI lists Codex in the CLI, SDK and IDE extension with no cloud features, billed at API pricing.

Current prices are on OpenAI's Codex pricing page. This guide does not repeat them because they change. In a Codex CLI session, /status shows your remaining limits.

What makes a Roblox session use more

• Each MCP server you add. OpenAI says every MCP server adds context to your messages and uses more of your limit. Keep only the servers you need while you work on your game.
• Large tool results. Walking the whole game tree of a big place or reading long scripts in full returns a lot of text. Ask for one folder, a depth limit or a line range instead: search_game_tree and script_read both support that.

Troubleshooting Codex and Studio

Roblox's own checklist is short: restart both Studio and the client, check that the command path exists, and check the config file's syntax. These are the cases specific to Codex.

Another route without MCP: RoCode

Codex with MCP suits people who are comfortable in a terminal or code editor. RoCode (RoCode.app) is a different route: an independent AI agent for Roblox Studio, with a web chat at rocode.app and a free plugin from the Creator Store that carries out the steps in the place you have open. It is not an MCP product and does not use Studio's MCP server, so there is no command line, config file or MCP server to set up. You pair the chat and the plugin once with the 8-character code shown in the plugin panel.

What to know before you switch. RoCode cannot use a ChatGPT or Codex plan. It runs its own choice of third-party models through OpenRouter: DeepSeek V4 Flash and GPT-OSS 120B on every plan, including the free Starter plan with 2 credits a day, GLM 5.3 and Claude Sonnet 5.5 from Pro, Claude Opus 5.5 from Studio and GPT-6 Astra from Ultra. With plugin 3.9 or newer, in Build mode, it can press Play for a short test of up to 30 seconds and read the errors, but it never moves a character or presses keys, and it does not take screenshots, run code directly, insert Creator Store models or generate meshes, all of which the Studio MCP tools can do. Its Ask mode answers questions about your place with read-only tools and changes nothing.

With plugin 3.9 or newer, RoCode can run a short playtest and read the errors, but it does not move a character or press keys, so press Play yourself and collect a few coins to confirm the fix. The prompt is an example written by us; the steps use the real tool names. How the plugin and pairing work. RoCode is an independent product, not affiliated with, endorsed by or sponsored by Roblox Corporation.

Questions

Sources and further reading

• Roblox Creator Docs: Connect to the Roblox Studio MCP server. the quick connect list with Codex CLI, enabling the server, the Windows and macOS commands, the tools and troubleshooting
• Roblox Creator Docs: Build with a coding harness. the first read-only check in Workspace, Script Sync with an agent
• Roblox Creator Docs: Script Sync. how saved file changes reach Studio
• Roblox Creator Docs: Version History. saving with notes and restoring a version
• OpenAI Help Center: Using Codex with your ChatGPT plan. plans that include Codex, the four clients, Codex Cloud, codex doctor and the untrusted policy
• OpenAI: Codex CLI. install commands, Sign in with ChatGPT and Git checkpoints
• OpenAI: Model Context Protocol. config.toml, shared configuration, codex mcp, timeouts, approval modes and ChatGPT on the web
• OpenAI: Permissions. what permission profiles govern, and that MCP servers use their own controls
• OpenAI: Agent approvals and security. Auto and read-only, protected .git, the flags, and automatic approval review of side-effecting MCP tool calls
• OpenAI: Permission modes. Approve for me sends eligible approval requests to automatic review
• OpenAI: Codex Cloud. cloud tasks started from ChatGPT on the web, mobile or the desktop app
• OpenAI: Custom instructions with AGENTS.md. when Codex reads AGENTS.md
• OpenAI: Developer commands. /init and /mcp verbose
• OpenAI: Codex pricing. plans, the API key option, /status and the cost of each MCP server
• OpenAI: Authentication. signing in with ChatGPT or an API key
• OpenAI: Codex MCP server removal. codex mcp-server is gone; outside MCP servers still work
• OpenAI: Terms of Use. minimum age
• Cursor Docs: Rules. Cursor reads AGENTS.md
• Claude Code Docs: How Claude remembers your project. Claude Code reads AGENTS.md