Zopaf

What zopaf MCP server does

The zopaf MCP server gives an AI agent a structured negotiation model instead of requiring the agent to calculate trade-offs in conversation. It represents negotiable terms, scores user preferences, accounts for a BATNA, and evaluates packages against the available trade space.

The engine is suitable for negotiations with several adjustable terms, including employment offers, vendor agreements, real estate deals, partnerships, term sheets, and settlements. The calling agent remains responsible for asking questions, presenting proposals, and explaining the result. Zopaf supplies the calculations and model state.

How it works

A negotiation begins with create_session, which returns the identifier required by subsequent calls. Add each term with add_issue, supplying options from least favorable to most favorable for the user. Numeric terms can instead receive a range through set_issue_range; this maps acceptable values to a 0–100 score.

Use record_preference to indicate which issues matter more. Each preference updates the issue-weight model. set_batna records the alternatives available if no agreement is reached, and the engine uses those alternatives to determine leverage.

After the model has enough information, generate_counteroffers creates three packages with equal user utility but different term combinations. The documented workflow presents all three at once. When the counterpart selects or resists parts of those packages, process_counterpart_response uses that reaction to estimate counterpart priorities and produce a second-round offer.

Tools and capabilities

The zopaf MCP server exposes these negotiation operations:

• create_session initializes a model and returns a session ID.
• add_issue adds a discrete term and its ordered options.
• set_issue_range scores a numeric term over an acceptable interval.
• record_preference updates relative issue weights.
• set_batna records alternatives and calculates leverage strength.
• generate_counteroffers produces three iso-utility packages.
• process_counterpart_response infers priorities and returns a round-two offer.
• analyze_deal scores a proposed package, reports captured value, and suggests trades.
• get_negotiation_state returns issues, weights, BATNA, frontier size, and the recommended next step.

The calculations use mixed-integer linear programming and combinatorial scoring. No LLM tokens are used for the optimization itself.

Setup and configuration

A hosted Streamable HTTP endpoint is documented at https://zopaf-mcp-production.up.railway.app/mcp. Configure that URL in an MCP client that supports Streamable HTTP. Claude Desktop uses a server entry with the endpoint URL and type set to streamable-http. Claude Code can register the same endpoint with claude mcp add zopaf --transport streamable-http.

The tool sequence is stateful: retain the session ID returned by create_session and pass it to the other session operations. No environment variables or local runtime configuration are specified in the provided setup instructions.

Limitations and notes

Zopaf is a calculation service, not a conversational negotiator. The connected agent must collect the user's priorities, decide how to communicate offers, and interpret the results for the user. Output quality depends on correctly ordered issue options, meaningful numeric ranges, recorded preferences, and an accurate BATNA.

The documentation describes three generated counteroffers as equal in utility for the user, not as identical packages. Counterpart priorities are inferred from package preference and stated pushback, so the response data supplied to the tool affects the resulting model.