DecisionMatrix MCP

A transparent, 100% deterministic Model Context Protocol (MCP) server that gives LLM agents a reliable multi-criteria decision analysis (MCDA) engine.

Agents are great at gathering options but unreliable at weighing them: they lose precision, apply inconsistent weights, and can't show their work. DecisionMatrix offloads the scoring to an exact, explainable engine. You provide options and weighted criteria (plus a score matrix); it returns a fully scored, ranked, and explained result — with per-criterion breakdowns, the methodology used, the weights applied, and a plain-language explanation.

Every number flows through decimal.js at 40-digit precision (never floats), so identical inputs always produce byte-identical output. The server is stateless — no database, no sessions.

🌐 Live hosted server (free, no install)

A public remote MCP server runs on Cloudflare's edge — point any Streamable-HTTP MCP client at it:

https://decisionmatrix-mcp.pages.dev/mcp

{ "mcpServers": { "decisionmatrix": {
    "type": "http", "url": "https://decisionmatrix-mcp.pages.dev/mcp" } } }

It runs in open mode on the free tier (no key, 15 calls/day per IP). Paid plans (Starter $12/mo · 5,000/day, Pro $39/mo · 50,000/day) are live via Stripe Checkout — buy a plan, get an API key instantly, and send it as X-API-Key. Self-host for unlimited calls with no keys. Landing page + pricing: https://decisionmatrix-mcp.pages.dev.

What it does

Six tools, all returning a uniform, agent-parseable envelope:

Scoring methods

Each criterion has a direction: benefit (higher is better — quality, speed) or cost (lower is better — price, latency, risk). Weights are relative; they are normalized to sum to 1 internally.

Consistent response envelope

Every successful response contains: status, method, winner, ranking (with per-criterion breakdown), methodology, weights_used, inputs_used, notes, and a natural-language explanation.

{
  "status": "success",
  "method": "weighted_sum",
  "winner": { "option": "Gamma", "score": 0.666667, "score_exact": "0.666667", "rank": 1, "tie": false, "tied_with": [] },
  "ranking": [
    { "rank": 1, "option": "Gamma", "score": 0.666667, "score_exact": "0.666667",
      "breakdown": [
        { "criterion": "Price", "direction": "cost", "weight": 0.5, "weight_raw": "3",
          "raw_score": "900", "normalized_score": 1, "weighted_contribution": 0.5 }
      ] }
  ],
  "methodology": {
    "method": "weighted_sum",
    "name": "Weighted Sum Model (Simple Additive Weighting)",
    "normalization": "min-max per criterion (best value -> 1, worst -> 0)",
    "score_range": "0 to 1 (higher is better)",
    "weighting": "Criteria weights are normalized to sum to 1; only their relative sizes matter.",
    "deterministic": true
  },
  "weights_used": [ { "criterion": "Price", "direction": "cost", "weight_input": "3", "weight_normalized": 0.5 } ],
  "inputs_used": { "options": ["Alpha","Beta","Gamma"], "method": "weighted_sum", "option_count": 3, "criterion_count": 3 },
  "notes": [ "Scores are normalized within this option set; they express relative standing, not an absolute grade." ],
  "explanation": "Using the Weighted Sum Model, 'Gamma' ranks #1 with a score of 0.666667, ahead of 'Alpha' (0.527778) by 26.32% ..."
}

Errors never cross the tool boundary as exceptions — they come back as a structured, actionable envelope:

{
  "status": "error",
  "error": {
    "type": "incomplete_scores",
    "message": "Missing 1 score(s) in the options x criteria matrix.",
    "hint": "Provide a score for every option and criterion. Missing: Beta / Weight."
  }
}

Design note — exact numbers: score is a deterministically-rounded number (6 dp) for easy consumption; score_exact / raw_score are full-precision strings so no precision is lost in JSON. Rankings are computed on the exact values, with input order as a stable tie-break.

Project structure

decisionmatrix-mcp/
├── worker-src/
│   ├── index.mjs        # Cloudflare Pages Function (_worker.js): MCP over Streamable HTTP + billing routes
│   ├── engine.mjs       # The deterministic MCDA engine: 3 methods + 6 tools + validation
│   └── billing.mjs      # Stripe Checkout + KV-backed API keys, quota metering, webhook
├── site/
│   ├── index.html       # Static landing / pricing / docs page
│   └── _worker.js        # Built bundle (esbuild output; git-ignored)
├── tests/
│   └── engine.test.mjs  # 21 core scoring-logic tests (node --test)
├── examples/
│   └── agent_example.mjs # End-to-end MCP client demo over HTTP
├── package.json         # build / deploy / dev / test scripts
├── wrangler.toml        # Cloudflare Pages config
├── .env.example         # Optional auth/rate-limit env reference
├── LICENSE              # MIT
└── README.md

Separation of concerns: engine.mjs is pure and transport-agnostic (import it directly in tests or any Node/Deno/edge runtime); index.mjs only handles the MCP JSON-RPC wiring, HTTP, CORS, and the auth/metering seam.

Requirements

• Node 18+ (for the build, tests, and local dev). Only two dev/runtime deps: decimal.js (math) and esbuild (bundler).
• A Cloudflare account (free tier is fine) to deploy the hosted version.

Run it locally

git clone <your-fork> decisionmatrix-mcp && cd decisionmatrix-mcp
npm install

# Run the test suite (no server needed)
npm test

# Serve the MCP endpoint locally via Wrangler (builds + runs Pages dev)
npm run dev          # -> http://127.0.0.1:8788/mcp

# Try the end-to-end client demo (hosted by default, or pass a local URL)
node examples/agent_example.mjs
node examples/agent_example.mjs http://127.0.0.1:8788

Quick manual call:

curl -s http://127.0.0.1:8788/mcp \
  -H 'content-type: application/json' \
  -H 'accept: application/json, text/event-stream' \
  -d '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{
        "name":"list_methods","arguments":{}}}'

Install via npm (stdio, no hosting)

Run the server locally over stdio with a single command — nothing to deploy:

npx -y decisionmatrix-mcp

Claude Desktop / any stdio MCP client (claude_desktop_config.json):

{ "mcpServers": { "decisionmatrix": { "command": "npx", "args": ["-y", "decisionmatrix-mcp"] } } }

This is the same deterministic engine as the hosted server, running on your machine.

Client configuration

Cursor — ~/.cursor/mcp.json

{ "mcpServers": { "decisionmatrix": {
    "url": "https://decisionmatrix-mcp.pages.dev/mcp" } } }

Claude Desktop — claude_desktop_config.json

Claude Desktop launches stdio servers, so bridge to the HTTP endpoint with mcp-remote:

{ "mcpServers": { "decisionmatrix": {
    "command": "npx", "args": ["-y", "mcp-remote", "https://decisionmatrix-mcp.pages.dev/mcp"] } } }

VS Code — .vscode/mcp.json

{ "servers": { "decisionmatrix": {
    "type": "http", "url": "https://decisionmatrix-mcp.pages.dev/mcp" } } }

Any Streamable-HTTP MCP client

Point it at https://decisionmatrix-mcp.pages.dev/mcp (or your self-hosted URL). If you enable auth, add X-API-Key (or Authorization: Bearer <key>) in the client's headers.

Tools & parameters

create_decision(options, criteria, scores, method="weighted_sum")

• options — array of names (["Vendor A","Vendor B"]) or objects ([{"name":"Vendor A","scores":{...}}]). Minimum 2, names unique.
• criteria — array of { "name", "weight" (>=0), "direction": "benefit"|"cost" }. At least one weight must be > 0.
• scores — the option×criterion matrix. Accepted shapes:
  • object map: { "Vendor A": { "Price": 100, "Quality": 8 }, ... }
  • array: [ { "option": "Vendor A", "scores": { ... } }, ... ]
  • inline on each option object.
• method — weighted_sum (default) · weighted_product · topsis (aliases like saw, wpm, ideal also resolve).

score_options(options, criteria, scores, method)

Same inputs as create_decision; returns the full scored matrix (per-option, per-criterion normalized scores + totals) without the winner narrative.

sensitivity_analysis(options, criteria, scores, method, variation=0.2, steps=10)

Sweeps each criterion's weight from -variation to +variation (fractional, e.g. 0.2 = ±20%) in steps increments (2–100), renormalizing the others, and recomputes the winner each time. Returns a robustness_score (share of scenarios the baseline winner stays #1), the fragile_criteria, and per-criterion flip points.