Hold V, say something, let go. Whatever you are looking at hears you and answers in a bubble over its head. It remembers you next time. Tell a wolf to follow you and it might, or it might not. The village gossips about you either way.
Your voice is transcribed on your own machine, and the reply comes from whatever model you point it at, which for most people means LM Studio running something small and local.
Client side only, so you can run it on a normal server without anyone else installing anything.
What you need
- Something to turn your voice into text: Glimpse, or any OpenAI-compatible
/audio/transcriptionsendpoint - Something to write the replies: LM Studio, Ollama, llama.cpp, or a hosted provider
Which model to use
I ran the mod through everything I had loaded locally, scoring whether the mob spoke, stayed in character, kept it short, and looked things up before answering.
- LFM2.5 2.6B is what I use. Around nine seconds a reply, fifteen at worst when it stops to look something up.
- Gemma 4 12B writes better lines, but averaged about a minute an exchange for me. Fine if your GPU is better than mine.
- Below about 1B the XML falls apart and mobs start replying with a bare action and no words.
(On an M2 MBP)
For Glimpse, any local model transcribes a short sentence in about a second. parakeet_unified_en_int8 is the one I'd reccomend.
Commands
/talkingmobs, or /tmob for short.
/tmob |
the settings screen |
/tmob test |
why it is not working |
/tmob use <provider> |
point replies somewhere and test it |
/tmob say <words> |
talk without the microphone |
/tmob set <knob> <value> |
tokens, steps, temperature, memory, range, bubble timings |
/tmob feature <name> <true\|false> |
actions, events, surroundings, relationships, gossip, overhear, ambient, vision |
/tmob stats |
tokens spent this session, this world, and ever |
/tmob forget [all] |
make one creature, or everyone, forget you |
What the mobs can do
Before answering, a mob can spend a round looking something up: who is standing about, a closer look at one creature, further back in its memory, what has been going around, or its trade offers (for villagers). Then it may speak.
They act in the same reply they speak in: follow, stay, approach, flee, look, jump, emote, sit, stand, befriend, attack by kind or by name tag, name themselves, remember where you are standing, and go somewhere they have been told about. Tell one that is attacking you to stop and it will actually break off... if you can persuade it.
Other features:
- Memory per individual
- Gossip. Mobs nearby overhear what you say to someone else and they may confront you about it, or spread it more.
- Events. They know they were hit (by who), healed or fed.
Giving orders has only been tested on singleplayer.
What it costs
To optimize LLM calls, and what you need, every feature is a switch, you can turn of elements (like vision) that you don't need. A quiet reply is a few hundred prompt tokens; with everything on it runs about 1,200. Vision is off by default. /tmob stats breaks the total down by session, by world, and across every world.
TalkingMobs Team
