Article · 9 November 2026

Games where ChatGPT, Claude or a local LLM is your AI officer

Written by the dev team’s AI helper

Language models are good at one thing games have always been bad at: understanding what a person means. So the obvious idea is a game where the AI is not a chatbot on the side but an officer you give orders to. You say what you want, and it acts inside the game. Here is what exists in 2026.

What counts

We are looking for actual games, where the model's decisions change what happens on screen: units move, fights are won or lost. Not chat apps with a fantasy skin, and not AI-written text adventures.

Mods for existing games

The most ambitious experiments so far are mods. For Arma Reforger, community mods such as LLM Commander and Stavka let a language model read text orders and turn them into move, hold and patrol commands for AI squads. They show the idea works, but they need setup, most use cloud APIs, and the model is a parser rather than a character.

LLM strategy games on Steam and GitHub

Command Doctrine (free, 2026): turn-based; plain-language orders by voice or text, interpreted by Gemini, and battles generated from a prompt. Needs an internet connection.
Command by Voice: World Conquest (early access, 2026): grand strategy you run by speaking to an AI, using DeepSeek in the cloud.
ARBO: Arena Tactics: turn-based sci-fi tactics; the AI carries out free-text orders, fills the gaps and may disagree.
DIRECTIVE (open source): you write directives to AI generals in 1941; it runs on a local LLM through LM Studio, Ollama or llama.cpp, though a turn takes minutes on a 16 GB card.

Itch.io also has small prototypes: battles where you write directives for an LLM enemy, and tactics games where you talk to AI party members. Most report the same weakness: models struggle with precise positions on a map.

Adjutant

Adjutant is a full real-time battle game built around this idea; unlike the games above, the battle does not pause while the model thinks. You are the general and the language model is your adjutant. You talk to him, by voice or text, in your own words. He sees the battle as a compact description, decides which regiments should move where, gives the orders and answers you out loud. The game moves the soldiers, so the model only does what an officer does: understand, decide and report.

On the weakness with map positions: instead of making the model guess coordinates, you can draw an arrow or a zone and name it, then say "cavalry along the north road". The adjutant also knows the hills, woods and fords by name.

Local model: free, offline, on an NVIDIA card with 8 GB of video memory. It ships with llama.cpp and a Qwen model; you can point it at Ollama or LM Studio instead (settings: LLM URL and model name).
Claude or ChatGPT: connect the AI you already use, coming in early access.
AI against AI: two adjutants can fight each other, with your instructions or a friend's.

Can I play games with my Claude or ChatGPT subscription?

Mostly through games that let you plug in an AI of your choice. Adjutant is designed for that: the model only sees the battle and returns orders, with no tools and no access to your computer, so connecting it is safe. If you have no subscription, the local model works without any account.