Tryll

On-device AI for games and tools. Tryll runs small language models — and speech recognition, speech synthesis, and vector search — on the player's own machine. Instead of a single prompt call, you compose an agent: a workflow graph of typed nodes that retrieves from your own data, calls your game's functions, speaks and listens, and reads live game state. First-party plugins for Unity 6 and Unreal Engine 5 drive it over a binary protocol. No cloud API, no per-token cost, no network round-trip.

Start here

New to Tryll? Read Tryll in Ten Minutes first — the whole idea in one sitting.

Then browse the Use Cases: talking NPCs, an ambient world, voice control, an in-game assistant. If you already have a game in mind, find the closest one and see which building blocks it takes. If you just want to make something with AI, skim them for an idea that stands out.

When you are ready to build, pick your engine:

• Unity 6

  Drop a TryllAgentComponent on a GameObject and chat with it.

  First inference in Unity →

• Unreal Engine 5

  Drop a UTryllAgentComponent on an actor and chat with it.

  First inference in Unreal →

What Tryll gives you

• A local model host. One server process, many clients. It loads a model once and shares it across every session that needs it, and it ships inside the Unity package and Unreal plugin — there is no separate server to install or point at. Model management →
• Agents are workflow graphs, not prompts. An agent is a typed graph of nodes — generate, retrieve, classify, branch, call a tool, speak — wired by named exits and passing values to each other through a per-turn slot blackboard. Workflows, graphs, and nodes →
• Your data, your rules. Ground answers in your own lore with on-device retrieval (dense, lexical, or hybrid), force machine-readable output with a GBNF grammar, and short-circuit off-topic or out-of-character input before a model ever runs.
• Voice both ways. Streaming text-to-speech with zero-shot voice cloning, and streaming speech-to-text with push-to-talk or hands-free turn taking — both on-device, alongside the language model.
• Live game state in, real actions out. Agent variables push the player's level, mood, or inventory straight into prompts and retrieval filters, and tool calling lets the model invoke your functions and answer from the results.
• Conversations that last. Older turns fold into a rolling summary, so a character remembers the start of a long talk at a fixed token cost. You can seed or rewrite history, and save an agent's state with your save game and restore it later. Keep long conversations coherent → · Save and load agent state →
• A whole cast on one GPU. Agents share a loaded model, each node's context window is sized on its own, and agents nobody is talking to can drop out of video memory until they are needed. Run many agents on a small GPU →
• Unity and Unreal plugins, with editor tooling. Idiomatic Unity and Unreal APIs, plus tools inside both editors: a visual graph editor, a model manager, a chat window to test an agent without entering Play mode, a turn inspector that shows what each node did, the Agent Log, and the Dialog Lab, which runs one scripted dialog across parameter variants and seeds, shows the results side by side, and also runs in CI.
• Models picked on evidence. Model Benchmarks measure each catalog model on Tryll's own jobs (speed, video memory, and how well it guards immersion), so you can choose one for your hardware before you download it.
• Built for coding agents. Every package ships an AGENTS.md, and the docs publish an llms.txt index, so Claude Code, Cursor, or Copilot can write your workflows and wire them into your game. Agentic coding with Tryll →

Explore the docs

• Getting Started — tutorials for your first inference in Unity and Unreal.
• Use Cases — what you can build, starting from what the player experiences, and which building blocks each one takes.
• How-to Guides — task recipes: RAG assistants, tool calls, streaming to UI, local model setup, guardrails.
• Concepts — the mental model, starting with Tryll in Ten Minutes: architecture, agents, workflows, RAG, dialog history.
• Reference — every field, every enum, every error code.
• Release Notes — what changed in each version, and how to migrate across breaking changes.
• Model Benchmarks — measured speed, VRAM, and guard quality per model, to help you pick one.

What you can build

• An NPC that talks in character — and keeps doing it over a long conversation, because old turns are trimmed to fit the context window for you. Design an NPC prompt →
• An NPC that knows your world — quest logs, item descriptions, faction lore, indexed on-device and retrieved per line of dialogue. Create a simple RAG assistant →
• An NPC that acts — the model calls your functions (open the door, set the price, start the quest), you run them, and it answers from the real result. Define and handle tool calls →
• A conversation you can hold out loud — the player speaks, the character answers in a cloned voice, all locally. Add voice output → · Use voice input →
• Output your game code can trust — a grammar makes a reply parse-safe by construction, and game code can own the numbers while the model owns the wording. Constrain output with a grammar → · Keep merchant prices deterministic →
• A character that stays in its world — reject jailbreaks and out-of-character prompts with scripted replies, before spending a model call. Build an immersion guard →

Get help

• Ask a question on the Tryll Discord. It is the fastest way to reach the team and other Tryll developers.
• Report a bug by following Report a bug. It says which logs and details each kind of bug needs, and where to post the report.
• Email works too, but it is the slowest route: [email protected].