Zopaf

Zopaf Negotiation Engine -- MCP Server

A negotiation math engine exposed as MCP tools that any AI agent can call. Zopaf computes Pareto frontiers, generates iso-utility counteroffers, and infers counterpart priorities from their reactions -- all through pure MILP optimization. Zero LLM tokens burned. The calling agent handles the conversation; Zopaf handles the math.

Quick Start

Claude Desktop

Add to your Claude Desktop configuration (claude_desktop_config.json):

{
  "mcpServers": {
    "zopaf": {
      "type": "streamable-http",
      "url": "https://zopaf-mcp-production.up.railway.app/mcp"
    }
  }
}

Claude Code

claude mcp add zopaf --transport streamable-http https://zopaf-mcp-production.up.railway.app/mcp

Generic MCP Client

Connect to the Streamable HTTP endpoint:

URL: https://zopaf-mcp-production.up.railway.app/mcp
Transport: Streamable HTTP

Tools Reference

How It Works

1. Create session -- Initialize a new negotiation model with create_session.

2. Add issues -- Define the terms on the table with add_issue. Each issue includes options ordered worst to best for the user (e.g., Salary: ['$150K', '$160K', '$170K', '$180K']).

3. Set ranges -- For numeric issues, call set_issue_range to map values onto a 0-100 scoring scale.

4. Record preferences -- Call record_preference as you learn what the user cares about. Each call updates the internal weight model.

5. Set BATNA -- Use set_batna to record alternatives. The number and quality determines leverage strength and anchoring strategy.

6. Generate 3 counteroffers -- Call generate_counteroffers to produce three packages that are equally good for the user but structured differently. Present ALL THREE simultaneously. Never lead with one and fall back to another.

7. Process counterpart response -- Call process_counterpart_response with which package they preferred and what they pushed back on. The engine infers their hidden priorities.

8. Get round-2 offer -- Returns a refined offer on the efficient frontier, with value split weighted by leverage.

Example: Job Offer Negotiation

create_session
-> {"session_id": "a1b2c3d4"}

add_issue(session_id="a1b2c3d4", issue_name="Salary", options=["$150K", "$160K", "$170K", "$180K"])
add_issue(session_id="a1b2c3d4", issue_name="Equity", options=["0.1%", "0.25%", "0.5%", "0.75%"])
add_issue(session_id="a1b2c3d4", issue_name="Signing Bonus", options=["$0", "$10K", "$20K", "$30K"])
add_issue(session_id="a1b2c3d4", issue_name="Remote Work", options=["On-site", "Hybrid", "Fully Remote"])

set_issue_range(issue_name="Salary", worst_acceptable=150000, best_hoped=180000,
    option_values={"$150K": 150000, "$160K": 160000, "$170K": 170000, "$180K": 180000})
-> {"scores": {"$150K": 0.0, "$160K": 33.3, "$170K": 66.7, "$180K": 100.0}}

record_preference(preferred_issues=["Salary", "Equity"], over_issues=["Signing Bonus", "Remote Work"])
-> {"learned_weights": {"Salary": 0.345, "Equity": 0.345, "Signing Bonus": 0.155, "Remote Work": 0.155}}

set_batna(alternatives=["Competing offer from Company B at $165K", "Stay in current role"])
-> {"leverage_strength": "strong"}

generate_counteroffers(target_satisfaction="ambitious")
-> {
  "counteroffers": [
    {"label": "A", "terms": {"Salary": "$180K", "Equity": "0.25%", "Signing Bonus": "$10K", "Remote Work": "On-site"}},
    {"label": "B", "terms": {"Salary": "$170K", "Equity": "0.5%", "Signing Bonus": "$0", "Remote Work": "Hybrid"}},
    {"label": "C", "terms": {"Salary": "$160K", "Equity": "0.75%", "Signing Bonus": "$20K", "Remote Work": "On-site"}}
  ]
}

process_counterpart_response(preferred_package="B", pushback_issues=["Equity"])
-> {
  "counterpart_priorities_inferred": {"Equity": 0.571, "Salary": 0.143, ...},
  "round_2_offer": {"Salary": "$180K", "Equity": "0.25%", "Signing Bonus": "$20K", "Remote Work": "Hybrid"},
  "value_split": "User gets 75% of surplus"
}

The engine inferred that the counterpart cares most about equity (57% of their weight). The round-2 offer concedes on equity -- where it costs the user less -- and captures value on salary and signing bonus. Both sides improve. The user captures 75% of the surplus based on their strong BATNA.

Use Cases

• Job offers -- Salary, equity, bonus, title, remote work, start date, PTO
• VC term sheets -- Valuation, board seats, liquidation preferences, anti-dilution, pro-rata rights
• Real estate -- Price, closing date, contingencies, repairs, inclusions, rent-back periods
• Vendor contracts -- Price, SLA guarantees, payment terms, exclusivity, renewal clauses
• Salary negotiations -- Base pay, bonus structure, review timeline, scope of role
• Business partnerships -- Revenue split, IP ownership, decision rights, exit clauses, territory
• Legal settlements -- Monetary terms, non-disclosure terms, admission of liability, timeline

Why Zero Tokens?

Zopaf is a math engine, not a language model. It runs MILP optimization and combinatorial scoring -- operations that are computationally cheap but tedious for an LLM to attempt in-context.

Your agent's LLM handles the conversation with the user, asks the right questions, and explains the strategy. Zopaf handles the optimization -- computing Pareto frontiers, generating iso-utility packages, solving preference weights from revealed choices, and positioning offers on the efficient frontier.

You bring the brain. Zopaf brings the calculator.