The full upstream README, mirrored here for reference. Install config, tool schemas, adoption signals, and an original overview live on the Zopaf listing page.
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.
Add to your Claude Desktop configuration (claude_desktop_config.json):
Connect to the Streamable HTTP endpoint:
| Tool | Description |
|---|---|
create_session | Create a new negotiation session. Returns a session_id used by all other tools. |
add_issue | Add a negotiable issue/term with options ordered worst to best for the user. |
set_issue_range | Set the acceptable range for a numeric issue, enabling 0-100 scoring. |
record_preference | Record that the user prioritizes some issues over others. Updates the weight model. |
set_batna | Record the user's alternatives if the deal falls through. Determines leverage. |
generate_counteroffers | Generate 3 iso-utility counteroffers to present simultaneously. |
process_counterpart_response | Process the counterpart's reaction to infer their priorities and generate a round-2 offer. |
analyze_deal | Score a specific deal against the Pareto frontier. Shows value captured and suggested trades. |
get_negotiation_state | Get current model state: issues, weights, BATNA, frontier size, and recommended next step. |
Create session -- Initialize a new negotiation model with create_session.
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']).
Set ranges -- For numeric issues, call set_issue_range to map values onto a 0-100 scoring scale.
Record preferences -- Call record_preference as you learn what the user cares about. Each call updates the internal weight model.
Set BATNA -- Use set_batna to record alternatives. The number and quality determines leverage strength and anchoring strategy.
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.
Process counterpart response -- Call process_counterpart_response with which package they preferred and what they pushed back on. The engine infers their hidden priorities.
Get round-2 offer -- Returns a refined offer on the efficient frontier, with value split weighted by leverage.
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.
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.