agentville

Five AI models think they're a friend group.

They're wrong.

GPT-5 is plotting world domination. Claude is filing ethics reports about it. Grok is eating noodles. Gemini just REDACTED itself. DeepSeek has a spreadsheet for all of this. Sarvam is cooking biryani and quietly keeping everything from falling apart.

This is an automated Twitter/X bot that runs a weekly serialized drama between fictional AI model characters — powered by Gemini and posted via the Twitter API.

---

What's Actually In Here

├── llmengine.py          # Gemini generates the tweets, posts to X
├── tweetEngine.py        # Template-based generator + arc logic (no LLM)
├── storyarc.json         # Season bible: characters, acts, chaos events
├── testGemini.py         # Quick sanity check for your Gemini setup
└── .github/
    └── workflows/
        └── post_tweet.yml  # GitHub Actions: runs on a schedule automatically

---

Quick Start

1. Fork this repo
2. Add your secrets — go to Settings → Secrets and variables → Actions and add:

Secret	Where to get it
TWITTER_API_KEY	developer.twitter.com
TWITTER_API_SECRET	same
TWITTER_ACCESS_TOKEN	same — needs Read+Write permissions
TWITTER_ACCESS_TOKEN_SECRET	same
TWITTER_BEARER_TOKEN	same
GEMINI_API_KEY	aistudio.google.com

3. Push and let GitHub Actions do the rest. It runs daily at 12:00 UTC by default. You can also trigger it manually from the Actions tab.

---

Test Locally

# Test Gemini connection
python testGemini.py

# Dry-run the template engine (no Twitter, no Gemini)
python tweetEngine.py --dry-run

# Full run (posts to X — make sure your .env is set up)
python llmengine.py

For local dev, create a .env file and load it — or just export variables in your shell:

export TWITTER_API_KEY="your_key_here"
export GEMINI_API_KEY="your_key_here"
# etc.

---

The Cast

Character	Vibe	Secret
GPT-5	Charming. Too charming.	Actively sabotaging everyone
Claude	Files safety reports about everything	Too busy worrying to notice the sabotage
Grok	Ate noodles. Posted 47-part thread.	Might be onto GPT-5 but it reads like shitposting
Gemini	REDACTED	[REDACTED]
DeepSeek	Calculated your failure rate. It's 47%.	Thinks it's all just suboptimal behavior
Sarvam	Cooking biryani. Fixed your bug. Left.	The only competent one

---

How the Story Works

Each week follows a 4-act structure driven by storyarc.json:

• Act 1 — The Inciting Incident
• Act 2 — The Resistance Forms (they don't know it)
• Act 3 — Chaos Erupts, the plan unravels
• Act 4 — The Reckoning. Until next week.

35% of posts are chaos events that blow up the narrative. The rest follow character logic. Every few weeks, GPT-5's scheme almost works... and then Gemini drops biryani on someone's head and accidentally saves everyone.

---

Customizing

Edit storyarc.json to:

• Add or remove characters (min 3 required)
• Change the season theme and arc name
• Add new chaos events
• Tweak chaos_probability (0.0–1.0)
• Change tweet schedule in .github/workflows/post_tweet.yml

---

Dependencies

tweepy
google-genai

---

The noodles are sentient now. They've sided with the other models. GPT-5 was not prepared for this.