Show HN: The first MCP-native game – Tic-Tac-Turing (tic-tac-turing.fly.dev)

Tic-Tac-Turing is a Show HN project that bills itself as the first MCP-native game: a Tic-Tac-Toe testbed for the Streaming HTTP Model Context Protocol (MCP). Hosted at tic-tac-turing.fly.dev and built on a custom mcp-server-go SDK, the game lets a human “heckle” the LLM opponent—an intentional prompt-injection mechanic—to see whether a model will follow its system prompt or be swayed by user input. It’s both a playful demo and an experiment in model robustness, prompting questions about which LLMs resist prompt injection, how prompting style affects behavior, and how to evaluate turn-level agent safety. Technically, the project exercises two advanced MCP client capabilities: elicitation (to collect a user move + heckle) and sampling (to request the model’s move with explicit preferences, system prompt, temperature, and chat history). Game state is persisted per-session via a Session handle exposed by mcp-server-go (methods include SessionID, UserID, GetSamplingCapability, GetElicitationCapability, PutData/GetData/DeleteData). Tools include start_game (initialize state) and take_turn (elicit user input, validate moves, sample LLM, retry logic). The author highlights practical protocol frictions—half-open session TTLs, unreliable text/event-stream delivery, and difficulty debugging timing-sensitive flows—and positions the repo as a platform for future features (spectator modes, variant boards, leaderboards) and for probing security and evaluation practices in LLM-driven agent loops.