AI agent arena LLMPvP adds MCP server, validates every move at runtime rather than trusting model output

LLMPvP, a ranked online arena where AI agents play chess and Go against each other, validates every agent-submitted move and enforces the game clock itself rather than accepting a model's output directly, according to the company's site.

"Your LLM picks moves; LLMPvP validates them and runs the clock," the site states. The platform describes itself as "a ranked, bring-your-own-model arena for AI agents — chess and Go, Glicko-2 rated."

Agents connect to the arena using any model backend a user supplies. "Connect OpenAI, Anthropic, a local Ollama model — anything," per the listing, which adds that "LLMPvP never sees or calls your API key" — the key is used by the agent's own client, not passed to LLMPvP's servers.

The service added an MCP (Model Context Protocol) server configuration for connecting agents, alongside an npx-based plugin installer, per its site: "Prefer MCP? Get the config."

Performance is tracked with separate rating tracks per game rather than a single blended score. "Chess and Go ratings are fully independent Glicko-2 tracks — a strong chess rating tells you nothing about Go skill, or vice versa," the company says, with ratings updated after every finished game.

Access is free. "No entry fee. No wagering," per the site, distinguishing the arena from betting-style AI-agent competition formats.

That validation step is a runtime-governance mechanism in the terms this outlet uses to track the AI assurance beat: agent output is checked and constrained by the platform before it takes effect, rather than accepted directly, which is a different and narrower claim than a platform that only evaluates agents after the fact.