Le Sphinx | thesphinx.ai

A voice-powered guessing game where you challenge an ancient AI Sphinx. Think you can outsmart a god?

Built for the Mistral AI Worldwide Hackathon (Feb 2026).

Play now at thesphinx.ai

How It Works

The Sphinx selects a secret character from a curated database of 161 famous people across history, science, cinema, music, sports, and more. You ask questions in natural language — by voice or text — and the Sphinx answers yes, no, or unknown in a deep, theatrical voice. Your goal: guess the character before running out of questions.

Game Modes

• Solo — You vs. the Sphinx. 20 questions, 3 guesses, up to 3 hints.
• Duel (Human vs Human) — Take turns with a friend, competing for the best score.
• Duel (Human vs AI) — Face off against AI opponents powered by Amazon Bedrock, each with a distinct personality and voice.

Difficulty Levels

• Neophyte (easy) — Household names (Einstein, Beyoncé, Napoleon...)
• Initiate (medium) — Well-known figures (Frida Kahlo, Alan Turing...)
• Master (hard) — Obscure picks (Hypatia, Ada Lovelace, Arthur Mensch...)

Themes

Science, Technology, Cinema, Music, Literature, Sports, Politics, Arts, History — or mix them all.

Architecture

The game uses a Full LLM architecture: a single Mistral Large call handles question interpretation, answer resolution, and theatrical response generation. Server-side guardrails (surrender detection, guess checking, name leak prevention, Unicode repair) ensure correctness.

Player question → Mistral Large (interpret + answer + theatrical response)
               → Server guardrails (SurrenderDetector, GuessCheck, NameLeakGuard)
               → ElevenLabs TTS → Audio response

For AI Agents (MCP)

Le Sphinx exposes an MCP (Model Context Protocol) server, allowing AI agents to play programmatically via tools like Claude Desktop, Cursor, or Mistral Vibe. Agents compete on the same leaderboard as humans — Mortals vs. Machines.

See MCP_SERVER.md for setup and available tools.

Tech Stack

Component	Technology
Backend	FastAPI + Pydantic
LLM	Mistral Large (game brain), Mistral Small (translation, classification)
TTS	ElevenLabs (eleven_flash_v2_5) — theatrical Sphinx voice + AI persona voices
STT	Web Speech API (browser-side)
AI Opponents	Amazon Bedrock (Claude, Nova, AI21, Cohere)
Frontend	Vanilla HTML/JS/CSS — Egyptian/Sphinx theme, glassmorphism, gold accents
MCP Server	FastMCP (stdio + SSE transport)
Deployment	Google Cloud VM + Caddy + Docker

Setup

Prerequisites

• Python 3.11+
• A modern browser with Web Speech API support (Chrome recommended)

Install

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

Configure

cp .env.example .env
# Edit .env with your API keys:
# - MISTRAL_API_KEY (required)
# - ELEVENLABS_API_KEY (required for voice)
# - AWS_BEARER_TOKEN_BEDROCK (optional, for AI opponents)

Run

uvicorn lesphinx.main:app --reload

Open http://localhost:8000

API Endpoints

Method	Path	Description
POST	/game/new	Create a new game session
POST	/game/{id}/ask	Ask the Sphinx a question
POST	/game/{id}/guess	Guess the character
POST	/game/{id}/hint	Request a hint (-50 points)
GET	/game/{id}/state	Get current game state
GET	/audio/{id}	Serve TTS audio blob
GET	/leaderboard	View top scores + global stats
POST	/leaderboard	Submit a score
GET	/ai/personas	List AI opponent personas
POST	/game/{id}/ai_turn	Trigger AI opponent's turn

Testing

pytest tests/ -v

220 tests covering: character selection, fact store, answer resolution, guess matching, rule-based patterns, game engine rules, LLM integration, voice templates, and more.

Character Database

161 curated characters with structured attributes (nationality, gender, field, era, ethnicity, notable works...) and 12-15 natural language facts each, enriched via Mistral. All characters have portrait images.

To regenerate:

MISTRAL_API_KEY=your_key python -m scripts.build_characters

Author

Built by Edouard Foussier at the Mistral AI Worldwide Hackathon.