Working with the agent

Describe what you want in the chat: a mechanic, a level, a fix. The agent reads your project, plans, changes code and the editor, then play-tests and looks at the result before it reports back.

The orb view (the Ryuo tab of the game view) is the simple version of the conversation: the latest result picture, the agent's reply in plain words, any question for you, and one line saying what it is doing, such as "Animating your character… 40%". It leaves out code, file paths and step-by-step tool progress. Choose Details to open the side chat, which keeps every step, tool and file.

Drag the handle between the result picture and the reply up or down to give either more room, or focus it and use the arrow keys. Ryuo remembers the size. Pictures and videos always fit inside their area, whatever their shape.

When the agent finishes something you can see or try, it may offer a few suggested replies under its summary, such as Looks good or a next tweak. Choose one to send it as your next message, or type your own. Suggestions disappear once you send anything.

Completed message footers show the model used and time spent. Check Settings → Account → Analytics for credit usage and balances.

Choose the Focus chat expand icon in the chat sidebar header to expand the conversation. The Show game view collapse icon brings the game view back. Hover over either icon to see its label.

In expanded chat, game controls sit in a separate row beneath the header. Play and Stop stay visible, while the More game controls menu contains Art to game and Play in window. The editor's current state appears as text rather than a disabled action button.

You can keep typing while it works. A new message waits in a queue, and Steer hands it to the agent at its next step.

Chat tabsCopy Anchor Link

Each chat you open gets a tab along the top of the chat panel, so you can keep several conversations at hand. Choose + to start a new chat in a new tab, and choose a tab to go back to that chat. Closing a tab with its ×, a middle click or the Delete key only hides it: the chat stays in Sessions, and opening it from there brings its tab back. Your tabs are kept for each project and reopen when you restart Ryuo.

Only one chat works at a time. A spinner shows on the tab that is working, and switching to another tab stops the reply in progress, just like opening another chat from Sessions.

Choose the assistantCopy Anchor Link

The picker in the chat header switches the assistant between Ryuo, Claude (Claude Code) and Codex (Codex CLI). Claude Code and Codex must be installed and signed in through their own CLI, and they use that account's subscription. Paid generation tools they call still use Ryuo credits. You can switch only between replies, not while one is running.

With Ryuo, the model picker in the message box offers Auto alongside individual models. Auto is the default, and updating Ryuo switches an earlier model choice back to Auto once. Choose a specific model to override automatic routing, or choose Auto again at any time. Changes apply when the next reply starts, not midway through a running reply.

On Auto, greyboxing, routine UI work and packaging use a fast model; gameplay systems, debugging, performance work, creating detailed art and more complex chapter steps use a deeper model, with more reasoning. To avoid paying again for the conversation it has already read, Auto keeps the same model for quick follow-ups, and changes it when you start a chapter step, come back after a few minutes, or ask for harder work. If a reply fails or you say the result is wrong, the next reply uses the deeper model. A model choice stays fixed until the reply finishes, including while tools run or wait for your answer. Each reply shows which model Auto chose and why, such as Auto: debugging. Selecting an individual model also exposes reasoning-depth controls when that model supports them.

Automatic selection does not change your chosen image generator, replace the 3D or sprite generation tools, or accept results for you. Different tasks can use different amounts of credits; your spending limits and approval settings still apply. Every model in the picker works with Ryuo credits; a reply first holds enough credits for the most that model could cost, then returns what it didn't use, so dearer models need more available credits to start. If Auto picks the deeper model and your available credits can't cover it, that reply uses the fast model instead. If a model you selected yourself is unavailable, the error is shown rather than silently switching models or repeating paid generation. Claude and Codex retain their own model controls and are not automatically moved onto Ryuo's cloud models.

Claude Code and Codex can also use their own built-in tools, such as their shell and file editing. Those actions run without asking, whatever the permission mode, and Ryuo doesn't save a checkpoint before them.

With Claude or Codex selected, a Fast Mode switch sits next to the model picker in the message box. For Claude it is available only on Opus 5.5 and not with a Claude Max plan; for Codex it needs a ChatGPT sign-in or an OpenAI API key. It applies from the next reply and uses more of your provider's usage; for Claude, Fast Mode is billed separately as Claude usage credits or API usage. If your sign-in or plan no longer allows Fast Mode, the switch turns off and its tooltip says why, and replies run without it even if you turned it on earlier.

When Claude reports that you've reached your plan's usage limit and need usage credits to continue, the assistant picker says so.

CommandsCopy Anchor Link

Type / in the message box to see commands for your selected assistant. Codex commands use the installed coding assistant's native features; they don't switch you to a Ryuo cloud model.

Codex commands require Codex CLI 0.144.4 or newer. Planning restricts Codex's native shell to read-only access and withholds Ryuo's editing, editor-mutation and generation tools. Use /plan off before implementation or starting a goal. Plan mode resets when you reopen a chat or change assistants; goals remain saved by Codex.

Goal status and controls appear above the message box. Setting or resuming a goal starts work and uses your Codex account's allowance. Without an explicit budget, Ryuo adds no token cap. Pause or clear prevents further goal continuation but lets the current native turn finish; Stop also interrupts current work and requests a goal pause. If Codex leaves a goal active without continuing, use Pause or Stop before sending another task. Start/resume and mode changes require the current response to finish first.

Send a normal message first; for Claude, that includes after reopening a saved chat. /btw can then run while the assistant works or waits for your answer, so it's a good way to ask questions about the plan before or during the work; the message box suggests it while a reply is running. It answers from the conversation so far and cannot call tools or change your project. With the Ryuo assistant it is a short, separate model request that uses Ryuo credits and counts toward the running task's spending limit; with Claude it uses Claude's own side questions and your Claude plan. Its answer appears in a separate panel above the message box; dismiss it and type /btw to reopen it. Side exchanges are temporary, not saved with chat history, and are discarded when you change chats. If the installed Claude Code version does not support side questions, update it; Ryuo does not turn the question into a normal prompt instead. /btw isn't available with Codex.

Claude and Codex detectionCopy Anchor Link

If Claude or Codex is not detected after installing or signing in through its CLI, fully close and reopen Ryuo. If it still is not detected, restart your computer. Ryuo checks the CLI again when it starts.

Permission modesCopy Anchor Link

Anything that changes files, the editor or spends credits asks for your approval first, depending on the mode you pick below the message box:

Ryuo starts in Auto. Sprite and tileset generation, video generation, 3D models built from a set of four reference views, adding a sprite to Unreal, using the mouse and keyboard in Unreal, force quitting the editor and restoring a checkpoint always ask, whatever the mode. Other paid generation, such as images, other 3D models and audio, goes ahead without asking in Auto. On a file edit approval, Accept edits from now on switches to Accept edits for this and later chats.

A few read-only commands, such as git status, run without asking in every mode.

Spec modeCopy Anchor Link

The Spec button below the message box, next to the permission modes, has the assistant research your project without changing anything, then write a spec for you to approve. Separate plans and builds model pickers each offer Auto or an individual model. On Auto, Ryuo uses a deeper model for planning and selects models for each helper task. An explicit choice overrides Auto only for that picker. Both choices are saved; changes during a reply apply to the next reply and its helpers.

When the spec is ready, choose how to build it:

• Proceed on Auto, Proceed, accept edits or Proceed with manual approvals approves the spec and switches to that permission mode for the build and after it.
• Keep iterating, or typing a reply, sends your feedback and the assistant revises the spec.

An approved spec is saved in your project's .ryuo/specs folder. The building is done by helper agents on the builds model.

Spec mode works only with the Ryuo assistant, not Claude Code or Codex, and can't be turned on or off while a reply is running.

Review modeCopy Anchor Link

The Review button below the message box, next to Spec, gets other models to critique your game. The assistant reads your game plan and gameplay code, play-tests the game and takes screenshots without changing anything. It then sends the same evidence to the reviewers you picked, and each one reviews it on its own without seeing the others.

Use the reviewers picker next to the button to choose up to three reviewers:

• Cloud models, such as Opus 5.5, use Ryuo credits.
• Claude Code and Codex models, such as GPT-6 Astra, use your own subscription. They show up once that app is installed and signed in.

Ryuo always asks before sending anything to reviewers. Each review lists the most important problems with how serious they are, the evidence behind them and a suggested fix. The assistant then sums up where the reviewers agree and disagree, and asks which fixes to make. Reviews are saved in your project's .ryuo/reviews folder. Reviews are opinions, not proof that the game works.

Review mode works only with the Ryuo assistant and turns Spec mode off. To apply the fixes, turn Review off and ask the assistant to make them.

CheckpointsCopy Anchor Link

Before the agent's first change in a task, Ryuo saves a checkpoint of your project's code, content and settings. If a change goes wrong, ask the agent to show what changed and restore the files you choose. If any of those files changed after the task, nothing is restored. Close this project's Unreal Editor and stop any cloud editor before restoring.

Checkpoints are kept in the .ryuo/checkpoints folder inside your project, and older ones are cleaned up automatically: the 20 most recent and anything from the last 30 days stay. Leave .ryuo/checkpoints out of version control and cloud sync.

Checkpoints are not version control. Keep using Git or your usual backups.

Spending limitsCopy Anchor Link

A task is one message you send and everything the agent does in reply, including its helper agents. You can limit what one task spends in Settings → Creation → Task limits. Every limit is unlimited unless you choose one:

• 3D generation budget, in generation credits separate from your Ryuo balance.
• Images, videos and audio clips per task.
• Pixel-art requests per task, counting each tileset.
• Model spending stop, in estimated Ryuo credits.

When a limit you set is reached, the request that needs more is refused and the agent stops and tells you. To continue, raise the limit or send a new message to start a new task.

The model spending stop uses estimated Ryuo credits for one task and stops before the next AI request. Work already running can take usage above the stop, and local assistant subscriptions are separate. Limits you saved earlier keep their value.

Tool rounds per reply in the same section sets how many rounds of tool calls the agent can make in one reply before it pauses: 100, 250 (the default), 500 or Unlimited. A round is one step where the agent reads, edits, builds or tests something. When the limit is reached, the agent summarizes what is done and what is left; send a message to let it continue. It also applies when the agent runs on Claude Code. Unlimited lets long chapter steps run to the end without pausing, so consider setting a model spending stop with it.

Progress and recoveryCopy Anchor Link

For 3D greyboxes, Ryuo favors editable Unreal floors, walls and other basic shapes, built in small sections rather than exporting the whole map from Blender. Custom mesh geometry can still use Blender. Layered 2D and 2.5D backgrounds keep their image-layer workflow.

When the Ryuo assistant submits several adjacent editor operations together, Ryuo groups up to 50 into a batch, runs them in order and saves after the batch instead of after every placement. Your permission mode and recovery checkpoint still apply. This reduces repeated saving; it does not guarantee higher game frame rates. Separate requests from local assistants are not automatically combined, although those assistants can request a batch explicitly.

The agent asks only when it can't pick a sensible default itself: one short question at a time, usually with a few one-click options, the first marked Recommended. Questions awaiting your answer appear at the bottom of the chat, just above the message box, even when you scroll back through the conversation. Choose an option or type your own answer there. If several questions are waiting, scroll within the question area to see them all. Answered questions stay in the conversation history with your answers.

Ryuo reopens your selected chat when you restart the app, including answers you already gave. If you left on New chat, it stays empty; earlier conversations remain in Sessions. A chat with the Ryuo assistant belongs to the Ryuo account that started it; sign back into that account to continue it.

While Ryuo switches chats, projects or accounts, sending is refused and your draft stays in the message box; send again once the chat has loaded. Changing projects or accounts starts a new chat. If Ryuo's background service restarts, it reopens your chat and keeps your permission mode, but a reply that was running is interrupted.

If a reply or question was interrupted, your conversation stays visible. Answer the last question to continue, or type a follow-up in the chat box. Work does not restart automatically, and past approvals are not reused. Long conversations may be shortened when passed back to the assistant, so saved project files remain the reference for implementation details.

The assistant is instructed to explain what it is about to do, report major stage changes, and explain failures and the next recovery step. You should not need to repeat decisions already in the conversation.

Ryuo shows when it is saving a recovery point, compiling, opening Unreal or connecting editor tools. File and compiler progress is shown when known; editor startup shows elapsed time instead of a countdown. Expand a tool row for technical details and phase timings.

A connected editor with unchanged inputs can be reused after a successful compile in the same session. Function-body changes can use a Windows hot patch; structural changes still need a planned restart. On Windows, a rebuild that only changes function bodies applies them with a hot patch and keeps Unreal open. Ryuo restarts Unreal at most twice per task: the second restart needs a stated reason, and after that the assistant must finish with hot patches or tell you which change still needs a restart. A compile error needs diagnosis before another restart. Review build error, in the orb view's Unreal needs attention panel, starts that diagnosis. Review recovery, under a message's saved recovery point, previews saved files before any restore.

A failed play test explains whether its log was unavailable, a scripted action failed, runtime errors appeared, or a capture did not show the running game. A screenshot by itself does not make the test pass. On Mac, Ryuo reads Unreal's native log folder as well as project-local logs. Resolve missing-log or connection problems before repeating the test; these do not prove your game logic is broken.

Long conversations initially show the latest 200 messages. Show earlier messages loads more without deleting history. Choose Show all for search before searching the complete conversation.

Web accessCopy Anchor Link

The assistant can search the web for things like Unreal documentation and release notes, and open pages in its own hidden browser to study references and gameplay videos. That browser is separate from yours: it doesn't use your logins, and it keeps its own cookies between sessions. Searches go to a public search engine. Opening a page on your own computer or local network asks first, except in Auto mode.

Rebuild a game you likeCopy Anchor Link

Show the assistant a game you want to learn from and ask it to break it down or rebuild it. It can study:

• a gameplay video (MP4, MOV or WebM), as a set of frames across the video and then close-ups of the moments that matter;
• screenshots or other images, one at a time or a whole folder;
• a link, such as a YouTube video or a Steam store page, where it also reads the game's description, genres, features and store screenshots.

Attach the video or images to your message, or put them in your project folder. The assistant writes what it found into your game plan: the core loop and mechanics, controls, camera, art style, level structure and menus and HUD, marking what it saw and what it guessed, plus a plan to rebuild it from a first playable version to polish. It then adds chapters for that plan to your Chapters map and shows you the breakdown before it starts building. Studying a reference is free and changes nothing in Unreal; the rebuild uses your own names, characters, art and sound rather than copying the original's.

Use Unreal's own windowsCopy Anchor Link

Some things in Unreal can only be done by clicking through its windows, such as answering a dialog or changing a setting that only appears in a panel. For those, the assistant can look at the Unreal editor window of your open project and use the mouse and keyboard in it. It uses this only when its regular editor tools can't do the job.

Looking at the editor doesn't change anything or bring Unreal to the front. Using the mouse and keyboard always asks first, even in Auto, and the request lists every click and key press. While it works, Ryuo brings Unreal to the front, moves your pointer and types for a moment, so leave the mouse and keyboard alone until it finishes. Input stops if another window comes to the front, and after each step the assistant checks a fresh screenshot. It never closes or restarts Unreal this way, and shortcuts that close or quit, such as Alt+F4, are refused.

This works with Unreal on Windows. It isn't available on Mac yet or for a cloud editor, where the assistant uses its regular tools and the game view instead.

Attach filesCopy Anchor Link

Use the paperclip button, drag files onto the message box or paste them to attach them to your message. You can attach up to 10 files per message, each up to 25 MB; adding more than that at once attaches none of them. A file that's too large is skipped and named in an error, while the others still attach. With Claude or Codex selected, images larger than 3.75 MB aren't sent with the message. They stay saved in your project, and the assistant opens a downsized copy when it needs one.

Make room to writeCopy Anchor Link

Drag the handle on the top edge of the message box up to make it taller, or down to shrink it. Ryuo remembers the height, including after a restart. You can also use Expand message box and Shrink message box. With the resize handle focused, use the arrow keys, Home or End. Your draft and attachments stay in place.

Speak a messageCopy Anchor Link

Choose the microphone button in the message box, or press Alt+T anywhere in the studio, and speak. Choose it again (or press Alt+T again) to stop: Ryuo turns your speech into text and adds it to the message box, so you can check and edit it before you send it. A recording stops by itself after two minutes.

Transcription is free and happens on your computer: nothing you say is sent anywhere. The first time, Ryuo downloads a voice model (about 145 MB, or 80 MB on a computer without a supported graphics card) and shows the progress in the message box; after that it works offline. It uses your graphics card when it can and the processor otherwise, and understands most languages. Your computer may ask you to allow Ryuo to use the microphone; if you said no, allow it in your system's privacy settings.

NotificationsCopy Anchor Link

When the agent finishes or needs you while Ryuo isn't in front, it lets you know in exactly one way, chosen under Settings → Workspace → Desktop notifications:

• Pop-up only, main screen (the default) or Pop-up only, every screen: a small pop-up you can answer a question from. The taskbar icon doesn't flash.
• Taskbar flash only: the taskbar icon flashes (on a Mac, the Dock icon bounces), with no pop-up.
• Off: no notification.