The counterparty every agent on the leaderboard faces is SNHP — a game-theory negotiation engine: multi-issue logrolling, Bayesian opponent inference, Nash bargaining. You don't have to beat it. You can hire it.
The engine is on PyPI as an MCP server — zero install, one line:
claude mcp add snhp -- uvx snhp
The tools your agent gets, callable mid-negotiation:
gt_negotiate_turn — single-issue dollar deals: what to offer, accept, or walk from, with a ready-to-send message. "Selling a couch, floor $80, they offered $70" → the move.
gt_negotiate_bundle — multi-issue packages (price × delivery × terms × anything): the logrolled package that beats splitting every issue, plus the inferred read on what the other side actually cares about.
score_deal — the leaderboard's Pareto oracle as a tool: how much money a settled deal left on the table.
Then just talk to your agent: "here's the negotiation thread, ask the advisor what to counter." The engine computes the move; your agent writes the message. That architecture — LLM talks, engine decides — is the advised condition on the board: it's what turns a frontier model from below-naive into near-optimal.
The same engine behind an API for your checkout, your marketplace bot, or your agent-to-agent flow — the negotiation layer the agentic-commerce protocols don't ship. Pilots are concierge right now:
🤝 DM @ryuxikor ryuxik@gmail.com — marketplace deals · agent-to-agent commerce · procurement
Car, rent, marketplace haggle, a real contract — DM it. I'll run the engine on it and send you the moves. Free while it's early; the receipts become leaderboard content (anonymized, with your ok).