_

Text adventures were the dawn of computer gaming. We're standing at a similar threshold. An agentic RPG engine where AI interprets your world's logic, systems execute with consequence, and narrative reflects reality.

Wishlist on Steam Join Discord Alpha
scroll
zork://adventure_1979
> You are standing in an open field west of a white house.
> There is a mailbox here.
> open mailbox
> Opening the small mailbox reveals a leaflet.
...
> The revolution begins where gaming began.
// THE VISION

From Text Adventures to Agentic Gaming

When Zork and Colossal Cave Adventure appeared in 1979, they invented a new medium. Players typed commands; the computer responded. The interface was primitive, but the concept was radical: the machine as a responsive storyteller.

That promise hit a ceiling. Every branching path required manual authoring. Graphics-driven games took a different path: rich interaction, but within tightly constrained systems. Open-world games feel free until you try something the designers didn't anticipate. Then the illusion breaks.

Large language models change what's possible—but only if they're harnessed correctly. An unconstrained LLM will happily invent outcomes, ignore established rules, and forget what happened three turns ago. That's not a game. That's collaborative fiction without stakes.

Synthasia takes a different approach: The LLM reasons within your world's logic, but the engine ensures that reasoning has consequences. Dice roll. States change. Actions matter. The AI is a participant in your game, not a god above it.

Text Adventures (1979)
Synthasia (2026)
// THE EXPERIENCE

What Does It Feel Like to Play?

Synthasia isn't just a technical achievement—it's designed to be genuinely immersive. Here's what players actually experience.

🌍

Any World You Can Imagine

Play in ancient Rome during the Republic. Navigate Cold War espionage in 1960s Berlin. Explore a cyberpunk megacorp dystopia. Wage war in a high-fantasy kingdom. Solve mysteries in Victorian London. The engine is genre-agnostic—worlds define their own tone, technology level, magic systems, and social structures.

Each world comes with its own visual skin, generated or crafted to match the atmosphere. A noir detective story gets moody shadows and sepia tones. A space opera gets sleek gradients and star-field backgrounds.

🎭

Your Character, Your Story

Create a character with the personality and capabilities you want to play. Define their background, their motivations, their speech patterns. The AI will remember and honor your choices—when you speak, your character speaks in your established voice. When you act, your options reflect who they are.

A charming con artist gets persuasion, deception, and fast-talk options. A battle-hardened mercenary gets tactical combat, intimidation, and survival instincts. A reclusive scholar gets investigation, lore insight, and research shortcuts. Same scene, different experience.

💬

Natural Conversations

Talk to NPCs naturally—not through dialogue wheels or keyword matching. Ask questions, make threats, flirt, bargain, confess, lie. The AI parses your intent, considers the NPC's personality and your relationship, and generates responses that feel alive.

NPCs remember what you've said. If you promised to help someone and didn't follow through, they'll notice. If you saved their life, they'll be grateful. If you betrayed them, they won't trust you again—or they might seek revenge.

⚔️

Tactical Combat

Combat isn't just clicking "attack." Describe your approach: "I want to feint left, then drive my blade toward their exposed side." The AI evaluates feasibility, considers your combat skills and their defenses, and generates options that reflect the tactical situation.

Environmental factors matter. Fighting in a narrow corridor limits movement. A muddy battlefield makes footing uncertain. Fighting at night changes visibility. The AI incorporates these elements into both your options and the narrative description.

📜

Meaningful Quests

Quests aren't just checklists—they're narrative threads that respond to your choices. Complete objectives through your preferred approach: combat, stealth, diplomacy, deception, or creative solutions the designers never anticipated.

The AI knows your active objectives and generates options that help you progress without railroading you toward a single solution.

⌚

A Living World

Time passes. NPCs have schedules. Events happen whether you're there or not. If you spend three days recovering from injuries, the world hasn't frozen—trade caravans have moved, political situations have evolved, NPCs have lived their lives.

Consequences propagate. If you help a village defend against bandits, they might spread word of your deeds. If you let the bandits raid, refugees might arrive at the next town with stories of the horror. Your actions ripple outward.

A Sample Play Session

synthasia://play_session
// THE SETUP
You're playing a disgraced noble turned spy in a Renaissance-inspired city-state. Your current objective: discover who's been leaking military secrets to a rival power. You've tracked the leaks to a merchant's guild.
// YOUR APPROACH
You could infiltrate the guild hall directly (Stealth), pose as a potential buyer to get inside (Deception), leverage your noble connections for a formal introduction (Persuasion), or confront a guild member in a dark alley (Intimidation). Each option appears because your character could plausibly attempt it.
// THE CHOICE
You choose deception. You fabricate credentials as a merchant from a distant city and approach the guild. The LLM generates the guild master's response based on their personality, the situation, and your performance. Your Deception check succeeds—they believe you.
// THE CONSEQUENCE
Inside, you notice a nervous clerk who seems eager to help. You could press him for information (Insight + Persuasion), follow him discretely (Stealth), or continue your merchant charade to get deeper access. Whatever you choose, dice will roll, and the narrative will reflect the actual outcome—not the one that makes the prettiest story.
// THE BREAKTHROUGH

AI Interprets. Systems Execute.
Narrative Reflects Reality.

"Agentic gaming" means the LLM functions as a semantic reasoning layer between your world's definitions and the engine's execution. It's not generating unconstrained text—it's making decisions inside a simulation.

synthasia://agent-runtime
// Player types: "I want to throw the vase at the chandelier"
// The AI doesn't just write a response. It processes:
 
Stage 1: Intent Classification
   → Complex action: environmental manipulation
 
Stage 2: Entity Extraction
   → vase (item), chandelier (feature), guards (NPCs)
 
Stage 3: Feasibility Evaluation
   → Dexterity check (moderate), Stealth check (easier if distraction succeeds)
 
Stage 4: Player Confirmation
   → Present interpreted action for approval
 
Stage 5: Execution
   → Dice roll. Consequences apply.
 
Stage 6: State Update via Tool Calls
   → Remove vase, update guard alertness, move player
 
Stage 7: Narrative Rendering
   → Prose reflects actual outcome, not desired outcome

Interpretation, Not Invention

The LLM understands your custom attributes, skills, and world rules semantically. When you create a character with "Affinity for Chocolate," the LLM knows to generate options where that trait matters—not because the engine has a "chocolate" subsystem, but because the LLM interprets what that attribute means.

Feasibility Before Fiction

When a player types a free-form action, the LLM evaluates whether it's possible given current context. It outputs a structured difficulty score. Then dice roll. Success or failure isn't a narrative choice—it's a system outcome.

Tool Use for State Changes

The LLM doesn't just describe outcomes in prose. It uses structured tool calls to modify game state: update relationships, move items, trigger events. Narrative follows state, not the reverse.

The Three-Layer Architecture

Layer 1: LLM Semantic Layer interprets, evaluates, proposes.
Layer 2: Engine Execution Layer rolls dice, validates, persists.
Layer 3: LLM Narrative Layer renders outcomes into prose.

The result: AI with genuine creativity, operating inside systems with genuine stakes. The LLM is your co-author in a story that could genuinely surprise you—not because the AI forgot the rules, but because the dice had their say.

// DYNAMIC GAMEPLAY

Same Scene. Different Character.
Different Game.

Because the LLM interprets your character semantically—not just as numbers on a sheet—your build changes what you see, what you can do, and what makes sense.

A guarded gate. A bored guard holds a clipboard. Your objective: get inside.

What the character notices:

The guard's attention drifts to their phone periodically. The service door has a faded "Authorized Personnel" sign. There's a delivery truck approaching in the distance.

Generated Options:
[Investigate] Intelligence

Study the guard's routine.

[Deceive] Deception

"Official inspection—you clearly weren't told about me."

[Wait] Patience

Time the approach with the delivery truck.

What the character notices:

The guard looks physically unimposing—no visible weapons. The gate's lock mechanism is exposed. The wall has footholds. No backup in sight.

Generated Options:
[Intimidate] Strength

"Open this gate or I'll open it through you."

[Athletics] Strength

Scale the wall while they're distracted.

[Attack] Combat

Just take them out—quick and quiet.

What the character notices:

The guard's uniform is slightly disheveled—long shift, probably tired. Coffee cup nearby. Photos on the desk—family? The service door's lock looks old.

Generated Options:
[Persuade] Charisma

"I'm delivering for the event—surely you can make an exception?"

[Bribe] Charisma

Slip 50 gold with a knowing look.

[Insight] Intelligence

Read the guard—what would make them cooperative?

Character creation — cyberpunk world with AI portrait and stat customization
💡
AI Interpretation in Action

The same prompt run with different character data produces materially different outputs. The scene hasn't changed—but the lens through which the character views it has.

// CORE SYSTEMS

Your Rules.
Engine's Execution.

The engine provides deterministic backbone systems that make AI-generated content meaningful. These systems run regardless of what the LLM proposes—providing stakes, consistency, and consequence.

Checks & Outcomes

The engine handles the math so the LLM can handle the meaning. D100 system for skill checks, D20 system for combat. Critical successes and failures emerge from the dice—not the narrative.

Combat

Turn-based combat with initiative, positioning, and tactical options. Creative actions are parsed, evaluated, and resolved. "I use my frying pan to reflect the fireball" gets evaluated for feasibility.

Dialogue & Relationships

Multi-NPC dialogue scenes where relationships evolve. The engine tracks relationship vectors, conversation history, and escalation states. NPCs won't suddenly trust you because the AI forgot your betrayal.

Quest Progression

Quests are integrated into generation context. When the LLM proposes actions, it knows active objectives, relevant NPCs, and what the player has already attempted.

Custom Player Actions

Players can always type free-form input instead of selecting generated options. The pipeline parses natural language into structured actions, evaluates feasibility, and resolves through the engine.

Turn & Time System

Actions take time. The world doesn't wait. Time advances, consequences ripple, and NPCs have their own schedules. The engine tracks temporal state so the LLM can generate time-appropriate content.

// ACTION SUCCESS CHANCES
Combat system — action selection with success percentages Multi-NPC dialogue scene — Cold War setting
// FEATURE: WORLD EDITOR

Writers Keep Control.
AI Accelerates.
Your Books Become Worlds.

The World Editor supports multiple creation philosophies—and they can be mixed freely in the same project. Build any world you can imagine, or transform existing stories into playable adventures.

// WRITE YOUR WORLD, YOUR WAY
✍️
Mode 01

Author Mode

Manual-first creation for writers who want precise control. Define every NPC's motivations, write every location description by hand, lock narrative fields from AI modification. Your prose, verbatim.

💡
Mode 02

Dreamer Mode

Prompt-first creation for rapid world scaffolding. Describe your world in a few sentences and let the AI expand it into a detailed, playable structure. Generate, refine, iterate.

📚
Mode 03

Upload Mode

Transform existing documents into playable worlds. Upload novels, lore bibles, history books—anything with narrative content becomes source material. Turn your favorite stories into games.

Upload a Book. Play Inside It.

How it works:

  1. Upload your document (any length—the system handles full novels)
  2. Automatic processing: The document is indexed, summarized, and entities are extracted
  3. Multi-stage generation: The world is built through dozens of specialized LLM calls
  4. Human refinement: Review, edit, and approve each element

Examples:

Upload The Count of Monte Cristo → Generate a revenge-conspiracy world in 19th-century France
Upload your cyberpunk world notes → Generate a corporate-dystopia with interconnected factions
Upload a history chapter → Generate an alternate-history scenario to explore
Upload your unfinished novel → Explore different paths your story could take

Your documents stay local. Nothing is sent to third parties beyond your configured LLM endpoints.

// PANEL CATEGORIES
01 Foundation Meta, LLM Directives, World Lore, Story
02 Structure Areas, Locations, Connections, Map Layout
03 Entities NPCs, Items, Weapons, Equipment Stats
04 Systems Loot Tables, Encounters, Story Flags
05 Character Player Stats, Relationships, Approaches
// AI ASSISTANT
synthasia://world-editor/ai
> Create a cyberpunk world on Mars
[OK] Generating world: "Neon Olympus" — 4 areas, 12 locations, 5 key NPCs
   Ares Industrial Zone... done
   Olympus Mons Underground... done
   Generating NPC: Jax "Chrome" Vasquez... done
> Make the tavern description more rundown
[MODIFY] Resolving entity: "Rusty Neon Cantina"
   Updating description... done
> _

Start With a Spark

One sentence becomes a full world. Areas, locations, NPCs, lore, item lists, and connections—all generated as a coherent foundation.

See Your World

A visual map shows every location, connection, and character. Drag nodes around, auto-layout with physics, and watch your world take shape.

Talk to Your World

Say "Make Elara smarter" or "Add a secret passage" and the AI finds the right entity and executes the change. Natural language world-building.

World Editor — full interface with AI Assistant, map visualization, and entity editing panels World Editor — complex world graph with interconnected locations and characters
🏪
Coming Soon

The World Store

Share your created worlds with the community—or sell them. The World Store will let creators publish their adventures, settings, and stories for other players to discover. Whether you want to give your world away for free or turn your creativity into something more, the platform will support both.

// CONTINUITY

Memory & Living World

Continuity across sessions. A world that doesn't forget.

Memory & Retrieval (RAG)

The engine maintains a semantic memory system that stores key events, NPC interactions, relationship changes, and quest progress. When generating new content, relevant memories are retrieved and included in the LLM's context.

No Narrative Amnesia

An NPC remembers if you helped them—or betrayed them. A location remembers what happened there. This reduces the common problem where AI-generated content forgets established facts.

Proactive Living World

The world doesn't freeze between player actions. Background processes pre-generate content for adjacent locations, track NPC activities off-screen, and let consequences propagate through connected systems.

Client-Side Vector Database

Vector databases running directly in your browser hold world lore, NPC memories, your uploaded documents, and novel chapters. Private by design, instant to query.

Quest Journal — tracking objectives across complex storylines
🔊
Word Gets Around

Your reputation travels. Help a merchant and her friends will greet you warmly next town over. Threaten a guard and the whole barracks hears about it before you arrive.

// ARCHITECTURE

Multi-LLM Orchestration

Director. Weaver. Clerk. And however many more you need. Each profile can be assigned to specific tasks, allowing precise control over quality, cost, and latency tradeoffs.

The Default Triad

Director High Reasoning
  • Action evaluation
  • Feasibility checks
  • Complex decisions
  • Strategic planning
Weaver Creative
  • Narrative generation
  • Dialogue
  • Atmospheric descriptions
  • Character voice
Clerk Fast Utility
  • Entity extraction
  • Classification
  • Simple parsing
  • Quick validation

Different tasks have different requirements. A feasibility evaluation needs strong reasoning. A narrative description needs creative flair. An entity extraction just needs speed. Routing each task to the appropriate model optimizes for quality, cost, and latency simultaneously.

// BEYOND TEXT

Content Pipelines

Visuals, voice, and novelization—enhancing the core gameplay loop.

🎨

Image Generation

Multi-provider image pipeline with priority scheduling. Generate character portraits, location art, or item illustrations. Use cloud providers, local ComfyUI workflows, or Stable Diffusion.

🎙️

Text-to-Speech

Streaming-capable TTS architecture with support for local Kokoro processing. NPCs can speak their dialogue in character-appropriate voices. WebGPU-powered, entirely on-device.

📖

Gameplay Novelization

Transform your play session into a chapter. The novelization pipeline takes your session's events and rewrites them as coherent narrative prose—a playable story becomes a readable story.

AI-generated location view — Cyberpunk neon district AI-generated location view — Ancient Rome with atmospheric prose
// UNDER THE HOOD

Built to Last

Everything runs on your machine. No accounts, no servers, no data leaving your computer unless you choose. The engine is designed to be modular, extensible, and completely in your control.

🤖

Any AI You Want

Use Anthropic, OpenAI, OpenRouter, or run models locally with Ollama. Swap providers any time without losing a single save file.

🔒

Fully Private

No accounts, no telemetry, no cloud dependency. Your story, your characters, and your worlds never leave your machine.

🧠

Three AI Tiers

A deep-reasoning model, a creative writing model, and a fast utility model—all working in concert. Right tool, right moment.

🎨

Multiple Image Engines

Cloud providers, local ComfyUI workflows with custom models, or Stable Diffusion. Your visual style, generated your way.

💾

In-Browser Memory

Vector databases running directly in your browser hold world lore, NPC memories, and uploaded documents. Private and instant.

🗣️

Real-Time Voice

WebGPU-powered text-to-speech narrates your story in natural, expressive speech—running entirely on your device.

// JOIN US

Building in the Open

Synthasia is an indie project exploring a new design space. We're developing in the open with community feedback as a core input—not an afterthought.

Roadmap

Now

Core loop stability, world editor refinement, feedback-driven iteration

Future

In-game World Store, creator publishing, community economy

How to Contribute

Wishlist on Steam

Visibility helps us continue development

Join Discord

Alpha testing, feedback channels, direct discussion

Report bugs and suggestions

We read everything. High-signal feedback shapes priorities.

Shape What Ships Next

We're building in the open. Every alpha report, balance note, and world-editor critique directly affects priorities. Come help us define what agentic gaming can become.

Wishlist on Steam Join Discord Alpha
Early Access Development