JevMC

A simple Java library for using Jev by TypeSafe.ai in PaperMC plugins.

JevMC is an unofficial Java SDK for the TypeSafe.ai API: a thin, dependency-free HTTP wrapper around POST /v1/systemone that exposes it as clean, typed Java classes. It contains no AI logic of its own.

Jev is a System One Model: you send it a state and typed questions, and it returns structured decisions with probabilities and confidence — no generated text to parse.

Features

• The 3 primitives: Noul (yes/no), Choice (pick an option) and Score (rate against a rubric)
• Asynchronous API via CompletableFuture, with automatic retries (exponential backoff) on 429/529 and network errors
• Zero shaded dependencies: uses java.net.http (Java 17+) and Gson, both already bundled with Paper
• Immutable types (records) and fully typed answers, including confidence

Installation

Add the library to your pom.xml (or install jevmc-1.0.0.jar into your local repository with mvn install):

<dependency>
    <groupId>dev.blancocl</groupId>
    <artifactId>jevmc</artifactId>
    <version>1.0.0</version>
</dependency>

Compile against Java 17 or higher (<maven.compiler.release>17</maven.compiler.release>). The jar also runs fine on newer JVMs (Java 21/25 servers included).

Basic usage

// Create the client (ideally a single instance in your plugin)
JevClient jev = new JevClient(System.getenv("TYPESAFE_API_KEY"));

// Noul: yes/no question -> NoulAnswer (0.0 = no, 1.0 = yes)
Noul toxic = new Noul("The message is toxic or insulting");

// Choice: pick an option -> ChoiceAnswer(choice, probabilities, confidence)
Choice action = new Choice("Suggested action for staff",
        Map.of("none", "The message is harmless",
               "warn", "Warn the player",
               "punish", "Mute or ban"));

// Score: rate against a rubric -> ScoreAnswer(score, legend, probabilities, confidence)
Score severity = new Score("Severity of the message",
        List.of("Harmless", "Annoying", "Offensive", "Severe"));

// A single call evaluates every question in parallel
jev.evaluateAsync(
        Map.of("player", "Steve", "message", "this server is garbage"),
        Map.of("toxic", toxic, "action", action, "severity", severity)
).thenAccept(response -> {
    if (response.get("toxic", NoulAnswer.class).isLikelyTrue(0.8)) {
        String suggestion = response.get("action", ChoiceAnswer.class).choice();
        // ...
    }
});

Usage in PaperMC

Golden rule: never call the API on the main server thread. Use evaluateAsync (events like AsyncChatEvent already run off the main thread) and hop back with Bukkit.getScheduler().runTask(plugin, ...) before touching the Bukkit API.

public class MyPlugin extends JavaPlugin {
    @Override
    public void onEnable() {
        JevClient jev = new JevClient(System.getenv("TYPESAFE_API_KEY"));
        Bukkit.getPluginManager().registerEvents(
                new ChatModerationListener(this, jev), this);
    }
}

A complete chat-moderation example that evaluates toxicity, severity and a suggested action in a single request ships in dev.blancocl.jevmc.paper.ChatModerationListener.

TypeSafe.ai API details

• Endpoint: POST https://api.typesafe.ai/v1/systemone
• Authentication: Authorization: Bearer <API_KEY> (get your key at console.typesafe.ai/keys)
• Model: jev-latest

Build and test

mvn package   # compiles and runs the tests