First Inference in Unity

Unpack a Tryll release, drop the package into your Unity project (the server is bundled inside it), configure one project setting, download a model through the Model Manager, and watch a single Generate node print its answer to the Console during Play Mode — no chat UI required.

Working with an AI coding agent?

The Unity package ships an AGENTS.md at its root. Attach it to your agent's context (Claude Code, Cursor, Copilot, Codex, …) before asking it to write Tryll code — it carries the graph/slot mental model and the ways Tryll differs from a cloud LLM SDK. Agents that fetch docs themselves can start from llms.txt.

What you'll build

A scene with a TryllAgentComponent on a GameObject that, on Start, connects to a locally-spawned Tryll server and creates a one-node agent. You trigger the first inference by right-clicking the component in the Inspector (Context Menu → Ask Tryll), and the answer streams into the Unity Console.

Before you start

• Unity 6 (6000.x) or later, with a blank 3D project.
• ~5 GB free disk space for the downloaded model and a working internet connection on first run.
• Familiarity with creating GameObjects, adding components, and writing a short C# MonoBehaviour.

You do not need to build anything from source — the distribution ships pre-generated FlatBuffers bindings.

---

Step 1 — Get the distribution

1. Download tryll-unity-<version>.7z from the releases page and extract it somewhere temporary (e.g. C:\Downloads\tryll\).

2. Inside is the Unity package — the server is bundled inside it, so there is nothing else to install:

   UnityPlugin/                       <-- the UPM package (everything you need)
   ├── package.json
   ├── Runtime/
   ├── Editor/
   └── .Server/Default/               <-- bundled server (rides inside the package)
       ├── tryll_server.exe
       ├── llama.dll, ggml*.dll, ggml-vulkan.dll
       └── data/
           ├── server-config.json
           ├── models.json
           ├── default-canned-responses.txt
           └── default-guardrail-patterns.txt
   

The archive ships no model weights — you download the models your graph references up front through the Model Manager (Step 4). CreateAgent fails fast if a referenced model is not already on disk; the server does not download it for you at agent-creation time. In Client Dev, downloaded weights land under <Project>/TryllStorage/models/.

---

Step 2 — Install the Unity package

Copy the UnityPlugin/ folder into your project's Packages/ directory and name it com.tryll.client. Unity auto-discovers embedded packages there, so there is no manifest.json edit to make:

<YourProject>/
├── Assets/
├── TryllStorage/                    <-- default storage root (outside Assets/)
├── Packages/
│   ├── manifest.json
│   └── com.tryll.client/        <-- copy UnityPlugin/* here
│       ├── package.json
│       ├── Runtime/
│       ├── Editor/
│       └── .Server/Default/     <-- server rides inside the package
└── <YourProject>.sln

Unity imports the package on focus; Window → Package Manager → In Project lists Tryll Client, and you should see no compile errors.

Accept the licence

After the import, the Welcome to Tryll window opens. You can also open it from Window → Tryll → Welcome. It has two setup steps:

1. Licence. Click Accept licence. The editor starts the server only after you accept the Tryll Runtime Licence. The licence includes telemetry, which is always on in the editor. The game you ship sends no telemetry. Telemetry opens the Telemetry page. Read the licence shows the full text from the package's LICENSE.md.
2. Server. This package bundles its server, so this step is done as soon as you accept.

The setup line reads 2 of 2 done. Withdraw takes the acceptance back and stops the editor's server.

Keep the package embedded here (a writable folder) rather than in a read-only cache. Model downloads and authored storage files live under <Project>/TryllStorage/, not inside the package.

Add this to .gitignore if you do not want models in source control (committing them is optional):

/TryllStorage/models/

Nothing else to copy

The server binary and its data/ ride inside the package's .Server/Default/ folder, so installing the package installs the server too — one folder to drop in, one to update on a new release.

---

Step 3 — Configure Tryll Client in Project Settings

Open Edit → Project Settings → Tryll Client and set:

Setting	Value	Why
Auto Launch Server	true (default)	TryllClient spawns the bundled server automatically. Uncheck to run a server yourself.
Llama.cpp backend	Vulkan (default)	Portable GPU API. Set Auto to prefer CUDA on NVIDIA (also ships CUDA DLLs in a player build).
Game Name	e.g. my-game	Human project label. Sent as the session name and as --project-name. Telemetry groups by a hidden project id; see Telemetry.
Telemetry Level	Basic (default)	Turns or Full send more while debugging. Takes effect the next time the editor launches the server.
Storage Data Folder	TryllStorage (default)	Project folder outside Assets/. The editor launches the server with --storage-root here.

There is no server-path setting — the location is resolved from the package automatically, identically in the Editor and in packaged builds. The editor sends content-free telemetry at the basic level. To send turn timings or conversation text while debugging, see Telemetry. Packaged players send nothing.

Settings are saved to Assets/Resources/TryllRuntimeSettings.asset (created on first access).

Models must be registered before use

CreateAgent fails fast if a referenced model is not already on disk — the server does not download it for you. Continue to Step 4 to download and register a model via the editor Model Manager before running the graph below.

---

Step 4 — Download and register a model

Model name fields in the Inspector are dropdown lists populated from the models you have registered in this project. On a fresh installation those dropdowns are empty — you must open the Model Manager, download a model, and register it before it appears.

1. Open Window → Tryll → Model Manager.
   The window launches the bundled server, connects, and lists every model the server knows about. Wait for the status indicator in the toolbar to show ● Ready.

2. In the model list, find a language model to start with.
   Recommended: Gemma 3 4B Instruct (Q4_K_M) (~3 GB download).
   Smaller / quicker: Llama 3.2 3B Instruct (Q4_K_M) (~2 GB download).

3. Click the model row to select it. In the Model properties panel on the right:

   • Status shows Absent — the files are not on disk yet.
   • Click Download. A progress bar appears and the toolbar shows active download count.

4. Once the status changes to Downloaded, set the Build registration to Experimental (usable in the editor, excluded from packaged builds) or Production (shipped in builds).

   • For this tutorial, Experimental is fine.

5. Close the Model Manager.

The Model Manager writes a TryllModelManifest asset to Assets/Resources/TryllModelManifest.asset. Any Model Name dropdown in the Inspector now lists the models you registered here.

Registration controls build inclusion

Experimental models are available during development but stripped by the build post-processor when you build a player. Production models are staged into the Server/ folder next to the player executable. A model that is downloaded but unregistered (status None) is invisible to the build system.

---

Step 5 — Create a workflow asset and set up the scene

1. In the Project window, right-click → Create → Tryll → Workflow Asset. Name it DemoWorkflow.

   

   The Create → Tryll submenu (the same menu as Assets → Create → Tryll).

   2. Select DemoWorkflow. In the Inspector, expand Graph: - Under Nodes, click + to add one entry. - Set Name = answer. - Click the Params type picker (shows <null> by default) and select Generate from the dropdown. The node's parameter fields expand below it — leave them all at their defaults; the Default Exit field stays empty, which routes the turn to END. - Set Start Node = answer. - Set Default Model Name using the dropdown — it now lists the models you registered in the Model Manager. Pick the one you downloaded in Step 4. 3. Create a new empty GameObject — call it TryllDemo. 4. Add a Tryll Agent Component (Add Component → Tryll → Tryll Agent Component). 5. Assign DemoWorkflow to the component's Workflow Asset field. 6. Un-check Auto Create On Connect — this tutorial wires the lifecycle explicitly so you control exactly when the agent is created. 7. Create a C# script TryllDemo.cs and attach it to the same GameObject:

using UnityEngine;
using Tryll.Client;

public class TryllDemo : MonoBehaviour
{
    private TryllAgentComponent _agent;

    void Start()
    {
        _agent = GetComponent<TryllAgentComponent>();

        _agent.OnAnswerText.AddListener((nodeName, text, isDelta, isFinal) =>
        {
            // Send = Streamed: every chunk is a delta, and the last one is an
            // empty terminator with isFinal = true.
            // Send = Whole: one chunk with isDelta = false and the full text.
            if (!string.IsNullOrEmpty(text))
                Debug.Log($"[Tryll] {(isDelta ? "delta" : "reply")} from {nodeName}: {text}");
            if (isFinal)
                Debug.Log($"[Tryll] {nodeName} finished its answer");
        });

        _agent.OnTurnComplete.AddListener((status, debugInfo, tokens) =>
            Debug.Log($"[Tryll] turn complete — status={status} tokens={tokens}"));

        _agent.OnError.AddListener(error =>
            Debug.LogError($"[Tryll] error: {error}"));

        _agent.OnAgentCreated.AddListener(() =>
            Debug.Log("[Tryll] agent ready — right-click this component and choose 'Ask Tryll'"));

        var client = TryllClient.Instance;
        client.CreateSessionComplete += err =>
        {
            if (!err.IsOk) { Debug.LogError($"[Tryll] CreateSession failed: {err}"); return; }
            _agent.CreateAgent();
        };

        client.StartSession();
    }

    [ContextMenu("Ask Tryll")]
    void AskTryll()
    {
        if (!_agent.HasAgent)
        {
            Debug.LogWarning("[Tryll] agent not ready yet");
            return;
        }
        _agent.SendMessage("In one sentence: what is Tryll?");
    }
}

---

Step 6 — Test without entering Play Mode

Before pressing Play, use the Chat Window to verify your setup is working:

1. Open Window → Tryll → Chat.
2. Confirm the Tryll Settings field shows your TryllRuntimeSettings asset (auto-found).
3. Drag the TryllDemo GameObject's Tryll Agent Component into the Agent Component slot.
4. Click Start — the window launches the server, connects, and creates the agent. The status label turns green: Ready.
5. Type a message and press Enter. You should see a streamed reply.
6. Click Stop when done.

See Test an Agent in the Editor for the full walkthrough including troubleshooting.

---

Step 7 — Enter Play Mode

Press Play.

On the first run you'll see:

1. A brief pause while the server starts and the client connects.
2. The model loads from disk (a few seconds the first time it is used in a session) — this only works because you already downloaded and registered it in Step 4; CreateAgent fails fast otherwise.
3. The Console prints: [Tryll] agent ready — right-click this component and choose 'Ask Tryll'.

Once the agent is ready:

1. In the Hierarchy, select the TryllDemo GameObject.
2. In the Inspector, right-click the TryllDemo script component header and choose Ask Tryll.
3. Watch the Console — answer deltas stream in as [Tryll] delta from answer: … lines, then [Tryll] answer finished its answer and [Tryll] turn complete.

---

Where to see what's happening

Open the Console window and filter by [Tryll]:

Prefix	Source	Useful for
[TryllClient]	TryllClient singleton	Connection attempts, session config, agent lifecycle.
[TryllAgent]	TryllAgentComponent	Agent created/destroyed, DownloadProgress, errors.
[TryllServer]	tryll_server.exe stdout (re-emitted by the client)	[Listener] Listening on 127.0.0.1:<port> (in the editor, auto-launch binds loopback on the Server Port project setting, default 9100, and falls back to a free port only if that port is taken; player builds always pick a free port), model load lines, server errors.

Client Dev writes one log file per run under %LOCALAPPDATA%\Tryll\projects\<project-key>\logs\ (project key from the sanitized game name plus a hash of the project directory).

---

What you built

• A session connected to a Tryll server the client spawned for you.
• One agent with a one-node graph authored in the DemoWorkflow Workflow Asset (Generate node, driven by a model you downloaded and registered in the Model Manager).
• Model output surfaced as OnAnswerText events, which stream token-by-token by default (Send = Streamed on the Generate node); set Send = Whole for a single full-reply frame instead of delta chunks.
• A [ContextMenu] method as the simplest way to trigger a turn from the Inspector without any UI setup.

---

Where to go next

• Test an Agent in the Editor — iterate on your workflow without entering Play Mode.
• How-to Guides — RAG, tool calls, guardrails, model management.
• Concepts — the mental model behind the graph abstraction.
• Unity Client API Reference — complete class-level reference.
• Manage models in the editor — download, register, and pin models with explicit progress.

---

Troubleshooting

• Compile errors on import — check that the package's Runtime/Generated/ folder contains .cs files. If it's empty, the distribution was packaged without the FlatBuffers step; contact support or re-download.
• "Accept the Tryll Runtime Licence in the Tryll Welcome window first." — the editor did not start the server because the licence is not accepted. The Welcome window opens. Click Accept licence, then try again (for example, click Start again in the Chat window).
• [TryllServer] server not found — the package's .Server/Default/ folder is missing or empty. A release package ships it populated; if you cloned the repo instead, build the Tryll server so the committed .Server/Default symlink resolves to your build.
• Model name dropdown is empty — open the Model Manager (Window → Tryll → Model Manager), download a model, and set its build registration to Experimental or Production. Unregistered models do not appear in Inspector dropdowns.
• Stuck at connecting — the server crashed or didn't start in time. Check the Console for [TryllServer] lines and the Client Dev log under %LOCALAPPDATA%\Tryll\projects\<project-key>\logs\ as a fallback.
• CreateSession fails — usually a server-side error reading models.json or initialising the inference engine. Check [TryllServer] Console output.
• OnError fires during CreateAgent — read the error message. Common causes: Default Model Name is empty or doesn't match any entry in models.json (fails fast — download it via the Model Manager first), or a route targets a node that doesn't exist.
• Right-clicking the script shows no Ask Tryll option — you must be in Play Mode with the agent already created. Check the Console for the agent ready log line first.
• Nothing appears in the Console — confirm OnAnswerText is wired in Start() and that the script is attached to the same GameObject as TryllAgentComponent. With the default Send = Streamed on the Generate node, OnAnswerText fires once per chunk (isDelta = true), the last call is an empty terminator (text is empty, isFinal = true), and OnTurnComplete follows. Set Send = Whole to get a single full-reply frame (isDelta = false, isFinal = true).