For Developers

In-game AI that works out of the box, in Unity and Unreal Engine

Basic NPC

3 topics

Natural dialogue

Characters understand plain language and answer in their own voice, with no dialogue trees to script.

Voice in and out

Speech-to-text and text-to-speech run on the player’s device, in many languages, without recording every line. The character starts talking before the reply is finished.

Roleplay guardrails

Players will try to break your characters. A three-step guard catches jailbreaks and out-of-world questions and answers them in character, before the reply model sees them.

Advanced NPC

5 topics

Knowledge and lore

Connect your world lore and quest data. Characters answer from what your world actually holds, and say so when they don’t know.

Talking is the gameplay

Flattery charms one character and annoys the next. Tryll reads each line for what your mechanic needs, such as charm, pressure or an offer, and your game decides what happens.

Player says“You have the kindest eyes in this whole tavern.”
Tryll scores the linecharm 0.8 · respect 0.6 · pushiness 0.1
Your game decidesAffection 42 → 47: she agrees to one dance
The character answers“Sweet words from a stranger… Fine. One dance.”

Emotions you can hear

Your game sets the mood, and the character carries it into what they say and how they sound. Each mood can have its own voice.

Actions and tools

Characters do things in your world: open doors, hand out quests, trade and follow the player. Your game checks every action before it happens.

Memory and summarization

Characters remember. Long conversations fold into a short summary, key facts go into a memory the character can search, and both save with your game.

Developer Tools

4 topics

Chat in the editor

Talk to your character inside Unity or Unreal, by text or by voice, without building the game or entering Play mode.

Visual workflows

Build a character’s logic as a graph in Unity and Unreal: look things up, classify, branch, act and speak.

Dialog Lab

Run one scripted conversation across prompts, models and seeds side by side, with checks on every reply. See which change really makes the character better.

Agent Log

Follow every agent while the game runs: what it heard, which path it took, how long each step took and what it said. Export the log to share a bug.

Performance & Memory usage

3 topics

Starts at 1.6 GB

A talking agent fits in 1.6 GB of graphics memory, with speech running on the CPU. Want smarter characters? Pick a bigger model: the model cards show what each one costs.

1.6 GBBonsai 8B 1-bit
2.9 GBQwen 3.5 4B
5.7 GBGemma 4 E4B
One agent, 4K context

Faster than speech

Small models write 100 to 250 tokens a second on a gaming GPU, many times faster than a character speaks. The voice starts after the first sentence.

256tok/s · Llama 3.2 3B
174tok/s · Qwen 3.5 4B
105tok/s · Ministral 3 8B
~3tok/s · People talking
Measured: 1K-token prompt, RTX 5070 Ti. People talk at about 150 words a minute.

Check it yourself

Open the Memory window to see how much RAM and graphics memory your models and agents take right now, in your own project.