Your AI
Choosing your AI
Director of Realms doesn't lock you into a model. Run local models on your own hardware, bring a key from the provider you already use, or mix the two, with a model for each job, live cost tracking and caps that stop spending before a request is sent.
Local or cloud
On your machine or in the cloud
Where a provider publishes its model list, DoR reads it live when you connect, so new models turn up without waiting for an update. Several local servers can be connected at once.
Local runtimes
- Ollama
- LM Studio
- Unsloth Studio
- LocalAI
- KoboldCpp
- Any OpenAI-compatible server
Images: ComfyUI, Automatic1111 / Forge, or a custom bridge. Voices: Piper and Coqui XTTS, or free browser voices.
Cloud providers
- OpenAI
- Google Gemini
- Anthropic Claude
- Mistral
- DeepSeek
- Fireworks
- Replicate
- xAI
- Cohere
- Together AI
- Groq
Image generation: OpenAI, Gemini, Stability AI, FLUX from Black Forest Labs, Replicate, Together AI, Fireworks and Recraft. Voices: OpenAI Speech and ElevenLabs.
A model for each job
A different model for each job
Each engine can use a different model. Pick one outright, or let automatic routing choose by a preset (Auto, Cheapest, Good Value, Balanced, Quality or Best available) and fall back to the next candidate if a call fails.
Measured, not promised These are starting points from DoR's own tests, not a guarantee for every model of the same size. Quality follows the model you choose.
Costs
Spending caps
Every paid call, whether text, vision, images or speech, is tracked as you play. Set a session cap and a daily cap, and when one is reached the next paid request is stopped before it leaves your machine.
- Illustration: $1 a session, $3 a day. Just an example, not a price, a default or a typical cost. The caps are on from the start, set in US dollars; change them to suit your table, or set one to zero to switch it off.
- A warning at 80%, before anything stops.
- Unknown prices aren't treated as free. A call without a known price is flagged as untracked, never booked as zero.
- Images and speech count too. A free local render costs nothing and isn't counted.
Past the 80% warning, so the GM has been told. At the cap, the next request just doesn't get sent.
An example, not a price or a default setting.
Privacy & security
Where your keys are kept
Never in the world
API keys and endpoint overrides are stored in the GM's browser, not in the shared Foundry world, so they're never sent to your players.
Only the GM calls out
AI requests come from the GM's client alone, and each key is only sent to the provider it belongs to.
Player chat, cleaned
Before player chat reaches any prompt, role markers, chat-template tokens and hidden commands are stripped out, so a message can't pretend to be the GM.
Model text, sanitised
Anything a model writes that becomes Foundry HTML passes an allowlist first: no scripts, no event handlers, only plain web links.
Off unless you say so
Hardware telemetry, feedback sharing and community map sharing are all off by default, and you see what would be sent before you agree.
Keep it on your machine
Point DoR at a local model and requests go to your own server, not a cloud provider. Campaign data stays in your world either way.
Questions
Questions people ask
Is it free to run?
Local models cost nothing per request, just your electricity. Cloud providers bill you directly at their own rates, and DoR tracks that spend as you play. Membership tiers don't include AI usage.
Do I need a powerful computer?
Not for cloud models: the heavy lifting happens at the provider. For local models, the hardware follows the model; the documented local image setup, for example, suggests a 12 GB VRAM-class GPU.
What leaves my machine?
Only what a feature needs, sent only to the providers or local servers the GM configured: prompts, player chat, character and campaign context, and for images the pixels being worked on. Each provider handles that data under its own terms, so choose yours with care and tell your players before their chat is used.
Do I need anything besides Foundry?
For cloud providers, a small helper that ships with DoR relays requests on your machine, because browsers can't call most AI services directly. Local servers are tried directly first, and the helper steps in if the browser blocks the request.
Interested?
Follow along while I build it
Still in Alpha, so things will change. Follow along for free on Patreon; the software itself comes later.