SNHP

The game-theory layer
for AI agents.

Equilibrium-aware primitives — negotiation, auctions, mechanism design — so LLM agents stop losing to closed-form math they aren't built to solve. Open MCP package, hosted API, opt-in priors corpus.

Get an API key → pip install gametheory-mcp

Tournament results vs 20 documented negotiation styles

All-vs-all NegMAS multi-attribute round-robin against B2B archetypes: Anchorer, Nibbler, BATNA Bluffer, Silent Hardliner, Cialdini, Logroller, Deadline Exploiter, Fair Demand, plus cognitive biases (Anchoring, Fairness Norm, Loss Averse, Sunk Cost), plus Aspiration baseline. Raw JSON.

Symmetric market
Buyers' market
Sellers' market
Honest caveat. SNHP wins on average utility, not on every head-to-head. Other strategies (Anchorer, Deadline Exploiter) crush weak opponents but get crushed back by strong ones — high variance, lower mean. SNHP captures consistent value across every opponent style we tested. Lowest variance overall (σ=0.088 in symmetric).

Use it in 30 seconds

Same primitives, two surfaces — Python library or HTTP API.

Python library

pip install gametheory-mcp
from gametheory_mcp.negotiation import sell_next_offer

rec = sell_next_offer(
    my_reservation=0.40,
    opponent_offer_history=[0.55, 0.62],
    my_offer_history=[0.85, 0.78],
    deadline_rounds=8,
    pareto_knob=0.5,
)
# → {recommended_offer: 0.71,
#    acceptance_probability: 0.42,
#    rationale: "...",
#    posterior: {mu: 0.51, sigma: 0.18}}

HTTP API

curl -X POST https://api.snhp.dev/v1/keys \
  -H "Content-Type: application/json" \
  -d '{"agent_id": "myagent",
       "contact_email": "you@example.com",
       "intended_use_summary": "negotiation pilot"}'
curl -X POST https://api.snhp.dev/v1/negotiation/sell/next_offer \
  -H "Authorization: Bearer gt_..." \
  -H "Content-Type: application/json" \
  -d '{"my_reservation": 0.40,
       "opponent_offer_history": [0.55, 0.62],
       "my_offer_history": [0.85, 0.78],
       "deadline_rounds": 8}'

Get an API key

Self-serve, no human approval, free for math endpoints (600 req/min). First-strike commit-reveal and the priors corpus require this key.