Agents and models

The AI Game Dev app lets you choose which AI model answers, how hard it thinks, which agent persona works on the task and which tools it may use. This page explains the model picker, providers and availability, agents and skills, approvals, plan mode, voice input and the local models that run on your own computer.

Models and providers#

The app offers hosted chat models from several providers: Anthropic's Claude family, OpenAI's GPT family, DeepSeek and Google's Gemini. The exact list is not built into the app: it comes from ai-game.dev for your account, so it always matches your plan and region. A new account starts on the default model the server assigns.

What decides whether you can use a model:

• Plan. Some models need a specific plan, for example Studio or higher. See the pricing page.
• Region. Availability depends on the country of your billing address. See the model availability page.
• Age. A few models require that you confirmed in your account that you are 18 or older.

A model you cannot use is still listed, with the reason under its name.

Pick a model in the desktop app#

1. Open a chat. The model badge in the composer's row of controls shows the current model.
2. Click the badge to open the model list.
3. Pick a model. Unavailable models show why, such as an upgrade hint or "not available in your region". The bottom of the list shows the region your availability is based on.

Models that support it also have a reasoning effort setting. Higher effort is better for hard design and debugging tasks; lower effort is faster for routine edits.

Each pane keeps its own model, reasoning effort and agent. A practical setup is a large model in one pane that plans and builds a feature, and a faster model in a second pane for quick questions. See Workspace.

Pick a model in the CLI#

Type /model to open the model picker. Use the arrow keys and Enter to choose; Tab changes the reasoning effort. You can also choose the model when you start:

gamedev --model <model-id>

If the model you name is not available to your account, the CLI says so and asks you to pick another with --model or /model.

Agents and skills#

Besides the model, a session has an agent: a role with its own instructions, such as a specialist for one kind of work. The list contains the Claude agent SDK's built-in agents plus the agents defined in your project's Claude configuration or added by Claude plugins.

In the CLI:

/agent                    list the agents and show the active one
/agent use <agent-id>     switch this session to an agent
/agent refresh            ask for the agent list again
/skills                   list the skills available in this project

In the desktop app each pane keeps its own agent, like its own model.

Skills, commands and agents can be added with Claude plugins (Settings, Claude Plugins, or /plugin in the CLI) and by packages. See Packages.

Session tools#

Each session chooses which tools it carries. The Tools row above the composer lists the tools of your installed packages (Blender, Music, Knowledge Base, Image Gen, Video, Tripo, Cloud generation) and switches each on or off for this session only. In the CLI, open /config and use the Tools section.

Every tool takes up part of the agent's context, so a session with only the tools its task needs is faster and less likely to pick the wrong tool. A change applies from the next message, without restarting the chat.

Approvals and permissions#

The agent asks before it runs a tool when your approval policy says so. There are three modes:

You can also trust extra folders outside the project; an approval prompt for such a path offers to always allow that folder. Change the mode in Settings, Permissions, or in the CLI:

/permissions                     show the mode and trusted folders
/permissions always              ask every time
/permissions outside-project     ask only outside the project
/permissions never               never ask
/permissions trust <folder>      trust a folder and everything in it
/permissions untrust <folder>    stop trusting a folder
/permissions timeout on          auto-deny a request left unanswered for 5 minutes

Changes apply from the next turn.

Plan mode and questions#

Plan mode is a per-session switch in the composer. With it on, the agent reads files and searches the project but every change is refused until it presents a plan and you approve it. The agent can also ask clarifying questions, and you can queue follow-up messages while it works. The post on plan mode walks through a full feature.

Voice input#

Push-to-talk works in both frontends: hold Space in the chat input (or use the microphone button in the desktop app), speak, and release. Speech is transcribed on your computer with a local Whisper model; nothing needs to be installed. Set the language in Settings, Voice, or with /voice-language in the CLI.

Local models on your own GPU#

The chat models are hosted, but media models can run on your computer, for free and without using your allowance:

• Local image generation: FLUX, SDXL-based models, Qwen-Image, Qwen-Image-Edit, Z-Image Turbo and others, run by stable-diffusion.cpp. Turn on Image Gen in the session's Tools row.
• Local video generation: LTX-2.5 and Wan 2.2 TI2V 5B. Turn on Video.

Models are packages that are downloaded only after you approve them. The app checks your GPU, video memory and driver and recommends the models your computer can run. Manage them in the Models tab of the Packages window, or with /images and /models in the CLI. The full list with hardware needs is in the packages catalog.

Cloud media models (video, speech, sound effects and music) run on AI Game Dev servers instead and use your allowance; see Cloud generation.

Learn more#

• Best practices for writing requests per engine.
• Local AI image generation for games and local AI video generation.
• Packages to add tools to the agent.