System requirements
The app itself is light. Almost everything a machine needs comes from who tells the story: a person or a hosted AI needs nothing special, and a local model is the one heavy part. Find your setup below.
The app
| Platform | Minimum | Notes |
|---|---|---|
| Windows | Windows 10 or 11, 64-bit, with 4 GB of memory | 8 GB of memory recommended. |
| macOS | Apple Silicon or Intel, with 4 GB of memory | A local model needs Apple Silicon. |
| Linux | 64-bit x86, with 4 GB of memory | AppImage, .deb, .rpm, Flatpak or tarball; Steam Deck included. The Flatpak cannot start a local model or an agent program. |
| Android | Android 8.0 or newer, 64-bit | A phone with less than 4 GB of memory starts with low effects on. A phone can host a world with a person or an OpenAI key telling the story. |
Measured on Linux, the desktop app uses about 1 GB of memory at the title screen and about 1.2 GB with its world open. The installer is close to 200 MB; a world with generated pictures takes a few hundred megabytes, and the optional sound library 76 MB more. An iPhone joins in Safari through an invite link; there is no iOS app yet.
Who tells the story
| Storyteller | What it needs | Where |
|---|---|---|
| A person | Nothing more. The app keeps the sheets, dice and maps. | Everywhere |
| An OpenAI key, or another compatible key | An internet connection. No graphics card; the model runs at the provider and bills your key. | Everywhere |
| An agent program (Codex, Grok Build, Claude Code, opencode) | The program installed and signed in on the computer that hosts. No graphics card; it runs on the program's own plan. Codex and Grok Build paint the pictures too. | Desktop app (not the Flatpak) and self-hosted servers |
| A local model | At least 16 GB of graphics memory. See below. | Desktop app and self-hosted servers |
With a key or an agent program the turn time is the provider's and the model's: pick a faster model there for faster turns. Each turn makes two or three calls (deciding the rolls and rules, then narrating).
A local model
| Graphics memory | Storyteller | Measured on | |
|---|---|---|---|
| Minimum | 16 GB (a 16 GB card, or 24 GB of memory on a Mac) | gpt-oss-20b, a 12 GB download | An AMD BC-250 with 16 GB of shared memory: 90 to 160 seconds a turn, playing the full rules |
| Recommended | About 28 GB for Q4, 45 GB for Q8 | Qwen3.6 35B-A3B, a 22 to 38 GB download | A Framework Desktop (AMD Ryzen AI Max+ 395, Radeon 8060S, 128 GB shared): Q8 took a median of 90 seconds a turn over 59 turns of real play, half of them between 59 and 127 seconds |
| Below the minimum | Under 16 GB, or no graphics card | Not supported |
Graphics memory is the card's own memory or, on a machine whose processor and graphics share one pool (Apple Silicon, AMD Ryzen AI Max), the part of that pool the system leaves the graphics: about four fifths, which is why a Mac needs 24 GB for the 16 GB minimum. The 35B model is a mixture of experts, so on a PC with a separate card its experts can wait in system memory at some cost in speed: half the system memory counts toward it, and a 16 GB card with 32 GB of system memory runs it at Q4. The desktop app, the server's guided setup and this page all reckon it that way.
A gaming PC with a 16 GB card: choose gpt-oss-20b. Cards such as the GeForce RTX 4060 Ti 16 GB, 4070 Ti Super, 4080, 5060 Ti 16 GB, 5070 Ti and 5080, or the Radeon RX 7600 XT, 7800 XT, 9060 XT 16 GB and 9070 series, hold OpenAI's open model entirely in their own memory, so turns stay quick and nothing leaves the computer. In the desktop app it is Story model for 16 GB under Local AI on this machine. With 32 GB or more of system memory the list may also mark the 35B at Q4 as fitting, even recommended, on the same card: it plays a little more reliably, at some cost in speed on every turn.
A smaller model can be made to run, and the server plays a model under 16 billion parameters on short rules by itself, but it skips rolls, attacks out of turn and narrates past the dice often enough that the table stops trusting it. Setups below the minimum are not supported, and problems on them are not treated as bugs. On a smaller machine, use a key, an agent program or a person.
No local model runs on an Intel Mac, on Windows or Linux on ARM, in the Flatpak build or on Android; those use a key, an agent program or a person. The settings and commands are in Local models; the desktop app installs and tunes one for you (Story AI on the desktop).
Local pictures
| Setup | What it needs |
|---|---|
| ComfyUI with an SDXL-class model | 8 GB of graphics memory or more. Newer models need more; the image model list shows each one's measured needs. |
| The desktop app's guided install | Windows or Linux, about 15 GB of disk with its Python packages, and Python 3.10 to 3.14 and Git installed first. Not offered on macOS. |
| Pictures from an OpenAI key | Nothing local. |
| Pictures from Codex or Grok Build | Nothing local: the agent that tells the story paints on its own plan, once one test picture succeeds. |
Pictures and a local model on one graphics card share its memory: add the two. A 16 GB card cannot hold gpt-oss-20b and an image model together, so take pictures from a key, from Codex or Grok Build, or from a second card. On a shared-memory machine with room for both, the server queues picture jobs behind each other, so a scene picture can arrive a little after its paragraph.
Hosting a group
- Players. A table seats up to eight players and the Dungeon Master. One machine can run several tables, but a local model narrates one turn at a time, so busy tables wait for each other.
- Upload speed. Voice is about 50 kbit/s for each listener of each person talking, and silence sends nothing; with six at the table that is about 0.25 Mbit/s for one voice and 1.5 Mbit/s if everyone talks at once. Pictures cost more: each scene picture or map is about 1.5 MB and goes to every player. About 5 Mbit/s of upload hands a new picture to five players in some 12 seconds; 1 Mbit/s carries voice and text alone.
- Where the voice goes. A world hosted in the app (desktop or phone) sends each person's voice straight to the others, so the host's connection carries only its own. A self-hosted server can relay every voice itself, and then its upload carries them all; see Voice chat setup.
- A self-hosted server for a person, a key or an agent program runs on any machine with Docker or Node and 2 GB of memory; see Server overview and requirements.