Assistant
Chat with your own AI model inside RBX Editor and RBX Debugger. It reads your script, looks through the attached game, and makes changes you can undo.
The assistant is a chat panel in RBX Editor (the Assistant tab on the editor's right edge) and in RBX Debugger (the Assistant view in the sidebar). It runs on a model you connect, with your own key, and it can work in your files and the attached game rather than only answer questions.
Connecting a model#
Open Settings (the cog at the top of any page, or Ctrl+,) and go to
Assistant → Providers.
| Option | What you need |
|---|---|
| Sign in with OpenRouter | An OpenRouter account. Opens in a new tab; no key to copy. One account covers Claude, GPT, Gemini, Grok and hundreds of others. |
| OpenAI, Anthropic, Google Gemini, xAI, DeepSeek, Mistral, Groq, Together, Fireworks, Cerebras, Perplexity, Azure OpenAI | An API key from that provider. Each has a Get a key link. |
| Ollama or Custom endpoint | Your own server's address. Anything that speaks the OpenAI chat API works: Ollama, vLLM, LM Studio, llama.cpp, LiteLLM. |
Paste a key and the model list loads straight away; that list is also the check that the key works. It keeps itself current: new models appear marked new, retired ones drop out, and if the model you were using is retired the assistant switches to its closest sibling and tells you.
Where your key goes
Keys are saved in this browser only. Each message passes through rbx.lol on its way to the provider, and nothing about it is stored there. The relay only reaches public addresses, so a self-hosted model has to be reachable from the internet (a VPS, or a tunnel), not only from your own PC.
More than one provider#
Connect as many providers as you like. The model menu at the top of the panel lists every model from each one, grouped by provider, with a search box once the list is long. Switch at any point, even halfway through a chat: the new model gets the whole conversation so far and carries on from it. Each reply is labelled with the model and provider that wrote it.
Long text and files#
Paste something long, like a whole script or a console log, and it goes in as an attachment above the message box instead of filling it. Attach files with the paperclip, or drop them onto the message box. Attachments stay with the message, so later questions in the same chat can still refer to them.
A reply runs for as long as the model needs. There is no time limit, however long it thinks or writes, and no cap on how many steps it takes. If a model repeats the same step and gets the same answer three times in a row, the reply stops there; say "continue" to let it try another way.
What it can do#
The assistant uses tools, and each one shows in its reply as it happens, for example Decompiling ReplicatedStorage.Modules.Combat or Searching Ember docs for "Slider".
| With no game attached | With a game attached |
|---|---|
| Read the file open in the editor and your saved files | Search the game's instances and read their properties |
| Edit your files | List the game's scripts and remotes |
| Search the Ember docs and look up Ember API names | Decompile a script and search decompiled sources |
| Look up Roblox classes and executor functions | Read the console and captured remote traffic |
| Run Luau in the game and read what it printed | |
| Read captured HTTP requests, send requests from the game | |
| Add, change and remove HTTP response mocks | |
| Fire a remote, or invoke one and read what it returned |
Remote and HTTP traffic are only there if capture is on in Remote Spy or HTTP Traffic; the assistant does not turn it on for you. Mocks it adds show in HTTP Traffic's mock list, where you can switch them off.
Every captured remote, request and packet has an Ask AI button: on the row when you hover it, in its detail dialog, and in its right-click menu. It opens the assistant with that item attached (a remote's latest calls as runnable Lua, a request's headers and bodies) and a question you can change before sending.
Changes you can take back#
Every edit the assistant makes is one step in the editor's undo history, so
Ctrl+Z and Ctrl+Y work as they do for your own typing. It also keeps a copy
of the file from before each edit: Restore beside the step in the chat puts
the file back, even after a reload or other edits since.
Before running code, firing a remote, sending a request or adding a mock in your game it shows what it will do and asks. Two switches in Settings → Assistant change that: Ask before running code in the game and Let it edit your files.
Thinking, effort and memory#
Thinking is whether the model reasons before it answers, and Effort is how much work goes into the whole reply: how long it thinks, how much it writes, how many tools it uses. Lower effort costs fewer tokens. The chips under the message box offer only what the selected model can do: a newer Claude shows Thinking: adaptive (it decides how much to think on each reply), a model that always reasons shows Thinking: always on (lower its effort to make either think less), one that never does shows No thinking, and the Effort chip steps through that model's own levels. Hover a chip for why. What each model can do comes from its provider and from models.dev, an open catalogue of models kept up to date, so new models work without an update here. Where a model thinks, the reply starts with a collapsible Thought for … line.
Memory is a short list of things the assistant keeps between chats, such as which executor you use or how you like answers. It can add to the list when you tell it something worth keeping, and every addition shows under the reply with Forget beside it. The full list is in Settings → Memory.
Chats are saved in this browser. Settings → Data clears them.
If a model refuses tools
Some models, especially small self-hosted ones, cannot use tools. The reply then says so and offers Chat without tools, which remembers that choice for that model. It can be undone under the provider's Advanced settings.