Core Concepts

A five-minute tour of the ideas the rest of the docs build on. If you've used a roleplay frontend before, most of this will be familiar — Pyre uses the standard ecosystem vocabulary.

The shape of Pyre

Pyre is local-first and BYOK (bring-your-own-key). There is no Pyre server: the app on your device talks directly to the AI provider you configured, and everything you make — characters, chats, lorebooks, presets, personas — lives on your device. That single fact shapes how everything else works: your library is yours, it works offline except for the model call itself, and it can sync between your own devices without a cloud in the middle.

The building blocks

Character

A character is who you talk to. Pyre uses the standard chara_card_v2 format, so a character carries a description, personality, scenario, first message, example dialogue, system prompt, tags, and more. You can import cards from the community, build them with the AI Creator, or edit every field by hand.

Persona

A persona is who you are in a scene — your name, a short description, an avatar, and optional voice samples. The model uses it to address you correctly and stay consistent. You can keep one default persona, switch per chat, or play with no persona at all. See Characters & personas.

Scenario (narrator card)

A scenario is a special kind of card: instead of one character, it's an omniscient narrator / Gamemaster that sets a world and voices a whole cast of NPCs. The AI Creator can build one for you, world and cast included.

Lorebook (World Info)

A lorebook is a set of keyword-triggered facts that get injected into the prompt only when they come up in conversation — so the model can "remember" a sprawling world without spending context on it constantly. Lorebooks can be bound to a character, a persona, or a single chat. See Lorebooks.

Preset

A preset bundles the prompt scaffolding (system prompt, post-history instructions) and optional sampling overrides (temperature, top-p, and friends) into one named thing you can switch in a tap. Pyre ships a tuned, sealed default and lets you fork it. See Presets & sampling.

Macros (template tokens)

Inside presets and prompts, tokens like {{char}}, {{user}}, and {{persona}} are filled in at send time with whatever character, persona, and lore are active — so one prompt adapts to every scene. The full list is in the Macros reference.

What happens when you send a message

When you hit send, Pyre assembles the prompt from the active preset, the character and persona, any lore that matched, and the recent conversation — substitutes the macros — and streams the reply back from your provider token by token. Along the way:

Variants & branching let you re-roll any reply or rewrite your own turn without destroying the timeline that followed it. See Chatting & roleplay.
Long-Term Memory keeps a running, branch-aware recap so long chats stay coherent past the model's context window. See Long-Term Memory.
Optional power tools — the Live Sheet (state tracking), the Script (story direction), and scene backgrounds — layer on top when you want them.

The provider

Pyre connects to any OpenAI-compatible provider. You bring the key; Pyre is the interface. It supports reasoning models, auto-detects model context windows, and offers a smart fallback to another provider if one refuses or fails. See Providers & connections.

Don't feel you must learn it all

A bundled example character and a connected provider are enough to start playing. Reach for lorebooks, presets, and memory when a scene actually needs them.