Getting Started¶
In about fifteen minutes you will install Tryll, download a model, and stream a reply from a language model running on your own machine. Pick the client that matches your project:
-
First inference in Unity Install the UPM package, add
TryllAgentComponentto a GameObject, see streaming text in the Console. The server is bundled and launches itself. -
First inference in Unreal ★ Hero tutorial. Install the plugin, add
UTryllAgentComponentto a sample actor, see streaming text in the Output Log. The server is bundled and launches itself. -
First inference in Python
pip install tryll-client, a 20-line script, streamed tokens on stdout. You point it at a server yourself — see Run the Tryll Server. -
First inference in C++ Build the client library with CMake and run the
test-chatconsole demo.
Each tutorial starts by downloading the distribution and ends with a working agent — including acquiring the model, which Tryll never does behind your back.
What your machine needs¶
Tryll runs the model on the player's hardware, so the hardware is part of the design. Check this before you start.
Operating system
- Windows 10 or 11, 64-bit. Tryll ships as Windows x64 binaries — the server, the Unity package, the Unreal plugin, and the C++ client library. There is no Linux or macOS distribution.
GPU
- 8 GB of VRAM or more is the comfortable target. That fits a quantised 7–8B model with a useful context window, and leaves room for speech synthesis and recognition running alongside it.
- 6 GB works if you stay around 3B models — a good fit for focused NPC dialogue, intent classification, and routing.
- More VRAM is mostly about freedom to experiment: larger models, longer contexts, and several agents resident at once. It is not required to ship something good.
- Speech-to-text and text-to-speech run on the CPU and cost no VRAM.
Disk and RAM
- 1.5–5 GB per language model, plus a working internet connection the first time you download one.
- Budget roughly the same again in system RAM: a model file is memory-mapped and offloaded to the GPU, so it counts in both columns.
Work out the real numbers for a specific line-up with Estimate the RAM and VRAM your setup needs.
Engine and language versions
| Path | Needs |
|---|---|
| Unity | Unity 6 (6000.x) or later |
| Unreal | Unreal Engine 5.7, a C++ project (Blueprint-only projects cannot compile the plugin) |
| Python | Python 3.10+ |
| C++ | Visual Studio 2022, CMake |
What you will build¶
Each tutorial ends at the same milestone: the user sends one
message, the server runs it through a minimal graph (just a
Generate node), and the client prints the streamed reply.
Everything else — retrieval, tool calls, voice, guardrails — builds on
top of this pattern.
After the tutorial¶
You have a one-node graph. These are the shortest paths to something that feels like a game feature:
- Author the graph visually instead of in code. Edit workflows in the Unity graph editor or in the Unreal graph editor
- Iterate without entering Play mode. Test an agent in the editor
- Write a character worth talking to. Design an NPC prompt
- Ground answers in your own lore. Create a simple RAG assistant
- Give the character a voice. Add voice output to an agent
- Feed it live game state. Drive prompts and retrieval from game state
If you want the why before the how, start with Architecture at a Glance.