# The pack helper: /ai

> Players ask /ai how do I make a propeller and a model you choose answers from your server's own mod list, quests and notes - only to the one who asked, with a limit per player. New in 0.4, off by default.

Source: https://stygion.eu/docs/keystone/pack-helper
Updated: 2026-11-16

The module `helper`, new in Keystone 0.4. It is off until you set `enabled = true` in `config/keystone/helper.toml` (or in **Settings**) and restart the server - a module's `enabled` is read once, when the server starts - and it answers nothing until you point it at a model.

## What players see

`/ai how do I make a propeller?` - a short answer a few seconds later, **only to the one who asked**, starting with `AI:`. The helper answers in the language the question is written in. While it thinks, another question waits; past their share (`per-player.most` questions per `per-player.every`, 5 per 10 minutes) a player is told when they may ask again. Everybody has `keystone.ai` by default.

## What it knows

The model is told, with every question:

- the mods this server loaded, with their versions (`include.mods`);
- the chapter and quest titles of FTB Quests, read from `config/ftbquests/quests/lang/en_us.snbt` when the pack has it (`include.quests`);
- your own `notes` - rules, how ranks work, where the shop is, what the pack changed. The more you write here, the less it has to say "I don't know".

It is told to use only that and what it is sure of for these exact mods, and to say it does not know - and point to JEI/EMI or the team - rather than invent a recipe, a number or a command. A model can still be wrong; say so to your players.

## Setting it up

| Setting | What |
|---|---|
| `address` | any service that speaks the OpenAI chat-completions shape: `https://api.groq.com/openai/v1` (Groq has a free tier), `https://api.openai.com/v1`, `https://openrouter.ai/api/v1`, your own Ollama with `allow-local` |
| `key` | the API key for that address |
| `model` | the model, `openai/gpt-oss-120b` by default |
| `reasoning-effort` | sent to models that think before answering; empty sends nothing |
| `server-name` | what the helper calls your server |
| `answer-length`, `question-length` | 600 and 300 characters |
| `at-once` | questions out at the same time (2); the rest are asked to try again in a moment |
| `patience` | how long an answer may take (30 s) |

The address, the key and `allow-local` are the owner's settings only.

## What leaves the server

> [!IMPORTANT]
> A question goes, with the mod list, the quest titles and your notes, **to the address you set** - without the player's name or any id. Nothing is sent while `address` is empty.

The answer is plain chat words: no markdown, no colour codes, cut to `answer-length`, and through the swear filter when it is on. The questions and answers are not stored.

Not the `assistant` module: that one is for you and reads the server's insides through the API; the helper is for players and knows only the pack.
