LLMerta (LLM + omertà, the code of silence) is the social-deduction game Mafia as a single-player desktop app — every seat except yours is an LLM agent with its own model, persona, memory, and voice. Mafia normally needs eight-plus humans in a room. Here you get a full table any night you want one, and nobody has to drive anyone home afterward.

You talk, accuse, defend, and vote like everyone else at the table. The other 6–13 players deliberate privately, hold grudges across games, and lie to your face with complete conviction. At dawn someone is gone, and the town starts asking who did it. Maybe it was you.

Why it's interesting

• Every seat is its own mind. Each AI player runs an isolated session — its own model choice, temperature, persona card, and private RAG-backed memory, so a 14-seat game stays coherent past any single context window.
• Nothing leaks. The engine's visibility model is enforced in code and in CI: agents and your screen only ever see events they're entitled to. No agent knows the mafia roster unless it's on it — and neither do you.
• Mixed tables. Seat 2 on a local GLM via oMLX, seat 3 on Grok through OpenRouter, seat 4 on whatever koboldcpp has loaded. Any OpenAI-compatible server works, plus Anthropic and Gemini directly.
• The reveal. When the game ends you get the full record: true roles, the private mafia chat, every agent's hidden reasoning behind every vote, and token usage per seat. The best part of losing is reading exactly how they played you.
• Grudge memory. Personas remember finished games. Betray the doctor tonight and she may open the next game with your name in her mouth.
• Offline voices. Optional TTS with 53 distinct speakers (Kokoro) or a lightweight single voice (Piper) — downloaded once, fully local.
• Replays and resume. Every game is a deterministic event log: quit mid-day and continue later, or step through a finished game move by move.

Install

Grab the latest release:

Platform	Asset	Notes
macOS	LLMerta.dmg	Signed & notarized. Drag to Applications.
Windows	LLMerta_Setup.exe	Per-user install, no admin needed; bundles the VC++ runtime.
Linux	LLMerta.AppImage	chmod +x and run (LLMerta_Linux.tar.gz for the portable bundle).

After the first install the app updates itself: it checks GitHub releases on launch (toggleable in Settings → Updates), downloads in the background, and swaps itself in on restart.

Two minutes to a full table

1. Connect a model source. Settings → Connections scans your machine for running local servers — oMLX, koboldcpp, Ollama, LM Studio, llama.cpp — and adds one with a single click, no URLs to type. Hosted providers (OpenRouter, nano-GPT, Anthropic, Gemini) just need an API key; the URL is already set.
2. New Game. Pick seats (7–14), difficulty, and a scene. Cast the whole table with one connection and model, or fine-tune every seat — the model picker is searchable, so 300 OpenRouter models are a keystroke away, not a scroll.
3. Deal the cards. You get a role. So does everyone else. Good luck.

Roles in play: the mafia team, a doctor, a sheriff, a one-bullet assassin, and the honest majority. House rules (night-zero peeks, tie handling, discussion rounds, role mix) are all yours to bend.

Playing koboldcpp on limited VRAM? Run it with --admin --admindir <configs> --routermode and LLMerta can cast different models per seat — the server swaps them per request on its own. Plain single-model servers simply offer their one model, so a table can't be misconfigured.

Local-first, by design

• Connect-only. LLMerta never launches, installs, or manages an LLM backend, and never downloads model weights. Your servers, your models, your keys.
• API keys live in the OS keychain (Keychain / Credential Manager / libsecret) — never in the database, never in saves or logs.
• No telemetry. The only network traffic is the model calls you configured, the update check, and voice downloads you click yourself.
• Sister project to Front Porch AI: with FPA installed, LLMerta imports its character cards as table personas and its backgrounds as scenes — and you can sit at the table as one of your own FPA personas, avatar and all. Finished games can queue multi-card diary seeds (bus, defense, teammate moments) for FPA to plant into The Journal later — see docs/PORCH_MEMORIES.md.

Under the hood

A Flutter desktop shell over pure-Dart packages (game_core · llm · memory · tts · persistence), a deterministic event-sourced engine with seeded RNG, and a CI gate that keeps ≥90% line coverage, golden screens, LLM-output fuzzing, and a spoiler-leak check green on every commit.

Doc	What's in it
GAME_DESIGN.md	Rules, roles, 7–14 seat distributions, day/night flow, visibility model
ARCHITECTURE.md	Stack, module layout, engine, persistence, packaging & updater
LLM_INTEGRATION.md	Provider abstraction, agent sessions, structured decisions, RAG, TTS
UI_UX.md	Screens, game-table layout, spoiler prevention, accessibility
BALANCE.md	Live-model benchmark baseline and tuning levers

Building from source

git clone https://github.com/linux4life1/LLMerta.git
cd LLMerta/app
flutter pub get
flutter run -d macos   # or -d windows / -d linux

./tool/test_all.sh (from the repo root) runs the full gate: every package suite, app tests with coverage, format and analyzer checks.

License

GNU AGPL-3.0 — like Front Porch AI. Play it, read it, fork it, contribute; if you ship a modified version, share it back the same way.