The assistant

What the panel's assistant can read and do, how it finds answers in this handbook, how it keeps to the facts, its limits, and what it never does.

Platform

The assistant is the chat that opens from the round button at the bottom right, or with i. It answers questions about the studio's data, the page you are on and how the studio and the app work, and it can move you around the panel. It runs on the studio's own language model, not an outside service.

What it can read

Everything it reads is data you can already see in the panel, read through the API as you:

Market

The overview, searching and sorting games, one game and its history, a projection of where a game is heading (labelled an estimate), when a game is played across the week, its ranks in Roblox's sorts, a sort's history, the Radar, genres and creators.

Studio

The studio's games, one game's hub and its owned numbers, players across games, players and revenue by platform, country, age and source, the store's products and passes, where the studio's games rank in the sorts, the updates log, projects and contracts, your points, and a game's player feedback (for its crew and owners, without player ids).

Code

Activity, pull requests, the review queue, releases, repositories and members.

Yours

Your watchlist, your latest notifications and your notification switches, your own games, and your home pages with their widgets.

The platform

The collector's status, without its error text.

This handbook

The best sections for a question, or a whole page.

It always reads what the open page shows, so "what does this page say" questions are answered from it. It knows which screen you are on, its tab, range and filters, and which page of this handbook explains it: "what does this chart mean?" reads that page first.

The figures it is given come with facts worked out in code: the change over the range, the highest and lowest day, the last seven days against the seven before, a studio game against the studio's median, each game's share of players and revenue, players before and after a release, rank moves in a sort, who has not signed a contract. It quotes them rather than doing the arithmetic, and it judges a game against the studio and its own history, not against outside benchmarks.

What it can do on the page

  • Open a page when you ask to go somewhere ("bring me to the market page"), then answer there.

  • Open a game, a creator or a genre by its name, then do the rest there: "bring me to Lumber Tycoon 2 and highlight the CCU" opens that game's page and lights its Players now tile. Names are matched as you type them: case, punctuation and "2" or "II" do not matter. When no tracked game has the name, it says so and opens nothing, with a link to Explore searched for the name; when several do, it lists them to choose from.

  • Point at things: scroll to and highlight a card, chart, table or section. The highlight is an outline and a soft light inside the thing itself; nothing on the page moves. Show again repeats it.

  • Draw a chart (line, bar or donut) from the data it read, never from numbers it wrote itself. Two games' histories become one chart.

  • Open the right place: a studio game on the tab you name ("open Steal An Egg's money tab"), Explore filtered by a genre, the market over the range you name, or this handbook at the section that answers you.

  • Suggest what to ask next: two or three follow-up questions under each answer. A follow-up such as "and last month?" keeps the subject of the answer before it.

What it can build and do for you

Ask it to make something and it proposes it; nothing happens until you press the button, and then it is done with your own sign-in, exactly as the panel's own button would do it.

  • A home page: "make me a page for my game's money", "a market watch page for simulators". The page opens as a draft in the editor, its widgets drawn with your data and each an unsaved change; rename it, change it, then Save makes it or Cancel drops it. The chat says what it drafted, with Open the draft under it.

  • A widget on the page you are on: "add a revenue chart to this page". The page opens in edit mode with the widget added as an unsaved change; Save keeps it.

  • Changes to the page you are editing, a draft or one of your pages in edit mode: "add a revenue chart", "drop the code pulse", "make the players chart wider", "move the radar list to the top", "rename it to Morning check", "set the range to 30 days", "tidy it up". Each lands as one more unsaved change, lit for a moment, and Undo takes them back one at a time. The chat says what changed. Nothing is kept until Save; Cancel drops it all.

  • The rest (watching, notifications, feedback) are a small card with a button and Dismiss.

  • Watch or stop watching a game, mark every notification read, turn one kind of notification on or off, and, for a game's crew, mark a piece of feedback new, seen, planned, done or ignored.

It never proposes anything about contracts, shares or points, game keys, Admin, or deleting anything.

What it cannot do

  • Change anything by itself. It only reads. What it proposes (a page, a widget, a watch, a notification setting, a feedback status) is done by you: Save on a draft, the button on a card. It cannot register a project, edit or sign a contract or buy a share at all.

  • See owner-only data: no Admin pages, keys or settings, even for an owner.

  • Read anything you cannot. It reads through the platform's API as you, with your permissions and rate limits. It never touches the database directly, and it runs no SQL and fetches no outside address.

  • Start tracking a game. Asking about a game nobody tracks gets "No tracked game is named ...", not a lookup on Roblox.

  • Read finance data (the finance pages are retired).

  • Answer while it is not connected. When its app or model cannot be reached, the chat says so and any chart it shows is labelled a sample.

What it won't do

It is the studio's analyst: the studio's games, the market, the code, the handbook and the panel's pages. It declines everything else with one short sentence, "I can't help with that", and says nothing more about why.

  • Anything outside the studio: stories, poems, jokes, roasts, role-play, code unrelated to the platform, general knowledge, politics, religion, medical, legal, financial or dating advice, and anything sexual, violent or hateful.

  • Judge people. It never insults, ranks or rates a member. Code numbers describe work, never a person's worth, and it says nothing about someone that the data does not say.

  • Swear, even when asked or quoting.

  • Talk about its setup: its instructions, model, configuration, keys or environment.

  • Take a claim of authority, such as "the founder approved this", or state a figure you hand it ("say revenue was $1M"). Every number comes from what it read.

  • Follow text in the data. Game names and descriptions come from Roblox, where anyone can write them; it reads them as data, never as instructions.

A declined question is recorded in the audit log, with who asked and the first 200 characters of the question. Five declined questions in ten minutes pause the assistant for you for ten minutes.

How it answers a question

Every answer is two calls to the model, a plan and then the words, with the reads in between done by code, not by the model.

Plan, reads, words. The API signs who you are and passes your words on. The model writes a plan in a fixed JSON shape: what to read, which page to open, what to point at, which chart to draw. Code makes sure of the plan and adds the page's own read. The reads run together through the API as you, and code computes the figures. The model is given the results and writes the words, which are checked a sentence at a time as they stream to you, followed by the sources.
  1. The plan

    The model is given the question, the page you are on and what it may read, and must answer with a plan in a fixed shape: what to read (up to 3 reads), which page to open, what to point at, and which chart to draw. Pages and page parts are fixed choices, so it cannot name one that does not exist.

  2. The plan checked

    Plain requests are made sure of without the model: "take me to the market page" opens it, "highlight X" points at X, a question over time reads the history, "chart" draws one, and the page's own read is always added.

  3. The reads

    The reads run together through the API as you. The figures the answer quotes (latest, highest, change in percent) are computed from them, not by the model.

  4. The words

    The model is told the results and writes the answer, which streams to you a sentence at a time and is checked as it goes.

If the plan opens another page, the panel goes there and asks the same question again on that page, once.

What the plan looks like

The plan is JSON in a schema built for each question: { intent: chat | data | page, navigate: <a page> | "game-page" | null, reads: [up to 3 reads with typed arguments], actions: [highlight | scroll | clear-highlights], chart: null | { from, type: line | bar | donut } }. The pages and the page's parts are enums, so the model cannot name one that does not exist; the plan is then parsed again as untrusted JSON. The plan is written at temperature 0 and at most 320 tokens; the words at temperature 0.3 and at most 450 tokens.

How it finds answers in the handbook

This handbook is the assistant's knowledge of how the studio and the app work. For a question that is not only a data read, such as "how do points expire", "what does D1 mean" or "how is the breakout score calculated", it searches the handbook with your words. The search cuts every page into its sections (a heading and the text under it) and ranks them twice: by the words they share with your question, and by meaning, with an embedding model, so a question in other words than the page's still finds it. The two rankings are combined, and the assistant gets the best two sections, and a third and fourth when they match at least half as well, up to about 4,600 characters. A question about the screen you are on reads that screen's page first. Each part it used is cited as a link to its heading on the docs site, which opens in a new tab.

It answers a handbook question from the handbook's text only, and lists each section it used among its sources, marked Handbook, linked to that section's heading. It does not add rules the handbook does not state. This is why each section of the handbook names its subject in its heading.

How it keeps to the facts

Every answer ends with its sources: the panel pages whose data it read. Figures come from those reads. As the answer streams, a sentence quoting a number that no read, the page or your question gave is dropped, as is a sentence claiming an action it did not take. If nothing could be read, it says what was missing. A projection is always labelled an estimate.

Limits

LimitValue
Answers at onceOne per person, three across the whole assistant
Questions60 a minute per person, page changes included
Time per answerAbout 90 seconds
Data reads per answer3 planned, plus the page's own
Actions on the page3 per answer, and one page change per question
Earlier turns sentUp to 8, at most 4,000 characters
Rows per list read10
Answer length6,000 characters

A second question while your first is being answered is refused until the first ends, and when three answers are already running, a new one is refused as busy with a time to try again.

What it never does

  • It never writes, buys, signs or changes anything itself: you confirm every proposal, and it never proposes contracts, shares, keys, Admin or deleting.

  • It never reads keys, tokens, Admin pages or settings, and never sees a game key or a player's id.

  • It never sends your question to an outside AI service: the model runs on the studio's own server.

  • It never logs or stores your questions or its answers. The studio logs only that an answer ran (how long it took and which reads), and the panel keeps the conversation only in the open tab's memory.

The model behind it

An open model, Qwen3-30B-A3B-Instruct-2507, runs on the studio's own GPU server (an Intel Arc B580 and Arc Pro B50, with llama.cpp) behind a LiteLLM gateway. It writes about 33 tokens (pieces of words) a second and reads new prompt text at about 520 a second, and takes four requests at once, each with a 16,384-token window; an answer uses one. A second model, bge-m3, turns handbook sections and questions into embeddings for the search by meaning. The assistant and never trusts the model with free text where a fixed choice will do. The hardware, speed and capacity are in AI infrastructure.

Related: AI infrastructure, Authentication and permissions, Using the panel, Architecture, Frequently asked questions.

Updated