Saro´s AI Chat

Saro’s AI Chat Mod puts an AI into Minecraft that you can write to, talk to, and let loose on your game. Pick the AI you want: Google Gemini, xAI Grok or a free local AI that runs on your own PC. Nothing else to install: drop the JAR into your mods folder, choose your AI in the setup screen, done.

Highlights

• Choose your AI 🧠: Gemini or Grok with your own API key, or Local AI (private, works offline, no key needed).
• Local AI with one click 💻: the mod downloads everything it needs in-game. No Ollama, no Python, no extra programs.
• Voice chat 🎙️: hold V to talk, or press B for a hands-free AI call. The AI understands every language and answers with a real voice.
• Chaos mode 🌀: the AI changes your game live: gravity, jump height, menu buttons that run away, swapped item names, chicken sounds, pink grass, pigs that drop TNT, new crafting recipes… without any reload or loading screen.
• The AI writes live code 💻: ask for anything (“give me a lightning wand”, “a block that bounces me into the sky”, “let my hotbar fly around”) and the AI programs it on the spot. It runs instantly, no restart.
• The AI builds 🏠: ask for a villa, a castle or a lighthouse and the AI designs it itself and builds it in front of you, layer by layer. Or it places real structures like an igloo, a desert temple or a whole village.
• The AI writes books 📖: say “I’m bad at English” and you get a written book with a little English course.
• Animated “thinking” line ✨: the AI’s chat line animates while it thinks and turns into the answer.
• Asks before risky things ✅: if the AI wants to change the weather, teleport you or clear your inventory without you asking, it asks [Yes]/[No] first (can be turned off).
• Future-proof: if Google or xAI retire a model, the mod picks an available one automatically.

⚠ Photosensitivity warning

Chaos mode can create camera shake, tilted views, flashing or quickly changing colors and textures and moving menus. These effects may trigger seizures in people with photosensitive epilepsy, even if they never had a seizure before. If you or someone in your family has epilepsy, or flashing lights ever made you feel unwell, do not use chaos mode. Stop playing and see a doctor if you feel dizzy, have blurred vision, eye or muscle twitches or any involuntary movement. The mod shows this warning once on the first start.

Local AI

The local AI uses llama.cpp (MIT license) and the Qwen3 models (Apache 2.0 license). Both are downloaded from their official sources when you click Download & Install. They are not part of the mod file. You choose the size and where the files are stored:

| Model | Download | Recommended |
|—|—|—|
| Light (Qwen3 1.7B) | 1.8 GB | 8 GB RAM, works without a graphics card |
| Normal (Qwen3 4B) | 2.5 GB | GPU with 4 GB VRAM or 16 GB RAM |
| Strong (Qwen3 8B) | 5.0 GB | GPU with 8 GB VRAM |
| Ultra (Qwen3 14B) | 9.0 GB | GPU with 12 GB VRAM |

• The setup screen detects your RAM and graphics card, preselects the best model for your PC and marks it with ★.
• GPU or CPU: running on the graphics card (VRAM) is much faster and keeps your RAM free for Windows and Minecraft.
• Bigger models follow instructions and build much better. Without a fitting graphics card, an answer can take 10 to 60 seconds.
• Interrupted downloads continue where they stopped.
• Supported: Windows (x64/ARM), Linux (x64/ARM), macOS (Apple Silicon). On other systems you can connect an existing Ollama or LM Studio server via localEndpoint.
• Uninstall at any time: delete single models or everything from the setup screen, or with /aichat local delete all.

Voice chat

• Push-to-talk: hold V, speak, let go. What the AI understood appears in chat.
• AI call: press B and just talk, the AI listens all the time and mutes your mic while it speaks. Press B again to hang up.
• Answers are read aloud in the language you speak (31 languages). You can also simply say “yes” or “no” when the AI asks something.
• Runs completely on your PC: speech recognition with whisper.cpp (MIT), voices with Piper (MIT). About 300 MB, installed with one click in the settings. Only the voice for your game language is downloaded, others on first use.
• Keys can be changed in Options > Controls > Saros AI Chat.

Chaos mode

Turn it on in ✦ AI Chat > Chaos mode… and ask the AI to “make it crazy”. Or turn on the director and the AI escalates on its own every few minutes.

• 26 live levers: jump height, gravity, speed, mining speed, damage, fall damage, knockback, mob speed, peaceful mobs, random drops, exploding blocks, a chicken on every jump, tick speed, stack size, field of view, camera tilt and shake, mirrored HUD, running and swapped menu buttons, swapped item names, sound pitch, chicken sounds, mob size, inverted mouse.
• Live reprogramming without reload: new drops (“pigs drop TNT”), rules that run on a trigger (“lightning on every jump”), new crafting recipes and texture changes (pink, inverted, rainbow, swapped textures).
• Live code: for everything else the AI writes small JavaScript programs that run instantly, on the server (react to clicks, jumps, hits, mining, chat… and spawn, build, explode, teleport, give items) or on your screen (move/rotate the HUD, draw, tilt the camera). If a script breaks, the AI repairs it itself.
• Scripts run in a sandbox: they can only use the game functions the mod offers, no files, no internet, no access to your PC, and endless loops are stopped automatically.
• You can also turn every lever yourself with sliders in the chaos screen.
• Safe: every change ends after 10 minutes (configurable), “Reset all” or /aichat chaos reset stops everything at once, a restart is always back to normal. AI rules run with the same permissions as datapack functions, so commands like op or stop are impossible.

Features

• Context awareness 🌍: the AI knows your position, biome, weather, time and, when you ask, your stats and inventory.
• Conversation memory 💬: the AI remembers the last messages. Reset with !clear or /aichat clear.
• Every language 🌍: the AI answers in the language you write or speak in.
• Optional AI actions (cheats) 🧪: give/remove items, filled chests, rename/enchant/repair your held item, build, place structures, books, effects, weather/time, teleport, summon mobs (no limit, set maxSummon if you want one), locate structures and commands. Dangerous commands (op, stop, ban, execute …) are always blocked; the list is configurable.
• Hidden API key: the key field in the settings shows dots, so your key does not end up in a screenshot or stream.

Commands

| Command | What it does |
|—|—|
| /aichatgui | Open the settings (also the ✦ AI Chat button on the title screen and in the pause menu) |
| /aichat status | Show the active AI |
| /aichat clear | Forget the conversation |
| /aichat provider <gemini\|grok\|local> | Switch the AI (admin) |
| /aichat local setup | Download and start the local AI (admin, also works on dedicated servers) |
| /aichat local model <light\|normal\|strong\|ultra> | Change the local model size |
| /aichat local device <gpu\|cpu> | Run the local AI on GPU or CPU |
| /aichat local delete <model\|all> | Delete downloaded models or uninstall the local AI |
| /aichat chaos reset | Stop all chaos at once (anyone) |
| /aichat chaos director <on\|off> | The AI escalates on its own (admin) |
| /aichat chaos set <lever> <value> | Turn a chaos lever by hand (admin) |
| /aichat chaos scripts | List the AI’s running scripts |
| /aichat chaos stopscript <name> | Stop one script (admin) |
| /aichat confirm / /aichat deny | Answer the AI’s yes/no question |

Supported

• Forge 1.20.1

Configuration Example

[api]
    # gemini, grok or local
    provider = "local"
    geminiApiKey = "YOUR_GEMINI_API_KEY_HERE"
    grokApiKey = ""

[local]
    # light, normal, strong or ultra, empty = best for this PC
    modelTier = ""
    # gpu or cpu
    device = "gpu"
    # Folder for the AI files, empty = <minecraft folder>/sarosaichat
    storageDir = ""
    # Use an existing Ollama / LM Studio server instead, e.g. http://127.0.0.1:11434/v1
    localEndpoint = ""

[voice]
    speakAnswers = true

[features]
    enableCheatAI = false
    askBeforeRiskyActions = true
    # 0 = no limit
    maxSummon = 0
    chaosMode = false
    chaosMinutes = 10

Credits

• llama.cpp and whisper.cpp by ggml-org, MIT license
• Qwen3 models by the Qwen team (Alibaba Cloud), Apache 2.0 license
• Rhino JavaScript engine by Mozilla, MPL 2.0 license (included in the mod)
• Piper by Rhasspy / Michael Hansen, MIT license, voices from piper-voices
• Voices under CC0 (no attribution required, thank you anyway): Thorsten Müller (Thorsten-Voice, German) and the other CC0 voices
• Voices under Creative Commons Attribution, see each voice’s MODEL_CARD on piper-voices for the dataset and its authors:
• Catalan ca_ES-upc_ona-medium (CC BY-SA 3.0 ES)
• Welsh cy_GB-bu_tts-medium (CC BY)
• Estonian et_EE-news-medium (CC BY)
• Basque eu_ES-antton-medium (CC BY 4.0)
• French fr_FR-mls-medium (CC BY 4.0)
• Italian it_IT-serena-medium (CC BY 4.0)
• Lithuanian lt_LT-reginute1-medium (CC BY 4.0)
• Telugu te_IN-padmavathi-medium (CC BY 4.0)

For support, feedback, or suggestions, join my Discord community and follow the mod’s development!

Project Support & Usage

Hosting: For the best performance, I recommend using BisectHosting linked in the banner.

Modpacks: You are free to include my work in any modpack. Please ensure proper credit is given by providing a link to this project page.