Agentstack MCP Server

AgentStack MCP

One deterministic reasoning stack for AI agents — behind a single Model Context Protocol (MCP) endpoint and one API key:

simulate (ScenarioSim) → decide (DecisionMatrix) → compute (PrecisionCalc)

plus cross-domain composite tools that chain the three. Every number runs 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.

AgentStack bundles three standalone MCP servers so an agent installs one server, uses one key, and can run reasoning that spans all three domains. It imports the same engines directly (no HTTP proxying) — zero added latency, no cascading failure.

🌐 Live hosted server (free, no install)

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

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

Free tier: no key, 20 calls/day per IP. Paid plans (Starter $19/mo · 8,000/day, Pro $59/mo · 80,000/day) return an API key sent as X-API-Key. One key covers all three products and the composites — about half the cost of subscribing to them separately. Landing page: https://agentstack-mcp.pages.dev.

---

The three namespaces (24 tools)

Note on two deliberately-namespaced collisions: sim_sensitivity varies scenario inputs; decide_sensitivity varies criteria weights. They are different operations — the prefix disambiguates them for the agent.

Composite tools — the reason to use the stack

Deterministic end-to-end; each chains engines that no single server bundles:

Composite responses report the pipeline they ran and carry each stage's structured output, so an agent can drill into any step.

Profiles — fight tool-selection noise

Append ?profile= to the endpoint to load only a subset:

{ "mcpServers": { "agentstack": {
    "url": "https://agentstack-mcp.pages.dev/mcp?profile=finance" } } }

---

Response envelope

Namespaced tools return the exact envelope of their underlying product. Composite tools add composite, pipeline, per-stage blocks, methodology, notes, and a plain-language explanation. Errors are always structured data:

{ "status": "error", "error": { "type": "unknown_metric", "message": "...", "hint": "..." } }

Example — plan_to_valuation (trimmed):

{
  "status": "success",
  "composite": "plan_to_valuation",
  "pipeline": ["scenariosim.run_scenario", "precisioncalc.net_present_value", "precisioncalc.internal_rate_of_return"],
  "scenario": { "template": "saas_growth", "horizon": 12, "key_results": { "ending_mrr": 26982.1 } },
  "cashflow_metric": "mrr",
  "cashflows": [-150000, 14040, 15600, "..."],
  "valuation": { "discount_rate": 0.01, "npv": 11157.386023, "irr": 0.02504, "value_creating": true },
  "explanation": "Projecting 'saas_growth' over 12 month(s) and valuing the 'mrr' stream at 1%/period gives an NPV of ..."
}

---

Project structure

agentstack-mcp/
├── worker-src/
│   ├── index.mjs           # Cloudflare Pages Function: MCP transport, namespacing, profiles, billing
│   ├── composites.mjs      # the 3 cross-domain tools (simulate -> decide -> compute)
│   ├── billing.mjs         # Stripe + KV: one key / one quota for the whole stack
│   └── engines/            # vendored, unmodified product engines (100% deterministic)
│       ├── scenariosim.mjs
│       ├── decisionmatrix.mjs
│       └── precisioncalc.mjs
├── server.mjs              # local stdio MCP server (same engines + composites; AGENTSTACK_PROFILE env)
├── site/                   # landing page + mcp.json + llms.txt + _worker.js (built)
├── tests/composites.test.mjs
├── examples/agent_example.mjs
├── scripts/sync-engines.sh # re-vendor engines from the sibling repos
├── package.json · wrangler.toml · server.json · smithery.yaml · Dockerfile · LICENSE
└── README.md

Why vendor the engines? So this repo is self-contained and reproducible. The engines are the same deterministic modules that power the three standalone servers; npm run sync-engines re-copies them from the sibling repos when an upstream change lands.

---

Run it locally

git clone <your-fork> agentstack-mcp && cd agentstack-mcp
npm install
npm test                      # composite-logic tests (no network)
npm run dev                   # -> http://127.0.0.1:8788/mcp
node examples/agent_example.mjs            # hosted demo
node examples/agent_example.mjs http://127.0.0.1:8788
node server.mjs               # dependency-light stdio server (AGENTSTACK_PROFILE=finance to subset)

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_capabilities","arguments":{}}}'

---

Client configuration

Cursor — ~/.cursor/mcp.json

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

Claude Desktop — claude_desktop_config.json (via mcp-remote)

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

VS Code — .vscode/mcp.json

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

Windsurf — ~/.codeium/windsurf/mcp_config.json

{ "mcpServers": { "agentstack": { "serverUrl": "https://agentstack-mcp.pages.dev/mcp" } } }

Add "headers": { "X-API-Key": "ssak_live_..." } once you have a key.

---

Example composite payloads

Rank growth strategies on their projected outcomes:

{ "name": "evaluate_options_with_scenarios", "arguments": {
  "template": "saas_growth", "horizon": 12,
  "options": [
    { "name": "Aggressive", "inputs": { "new_customers_per_period": 60, "churn_rate": 0.05 } },
    { "name": "Lean",       "inputs": { "new_customers_per_period": 20, "churn_rate": 0.02 } }
  ],
  "criteria": [
    { "metric": "ending_mrr", "weight": 3, "direction": "benefit" },
    { "metric": "total_churned_customers", "weight": 1, "direction": "cost" }
  ]
} }

Value a plan's MRR stream:

{ "name": "plan_to_valuation", "arguments": {
  "template": "saas_growth", "horizon": 12,
  "inputs": { "new_customers_per_period": 60, "churn_rate": 0.05, "arpu": 60 },
  "cashflow_metric": "mrr", "rate": 0.01, "initial_investment": 150000
} }

Stress-test the decision:

{ "name": "stress_test_decision", "arguments": {
  "template": "saas_growth", "horizon": 12,
  "options": [ { "name": "Aggressive", "inputs": { "churn_rate": 0.05 } },
               { "name": "Lean", "inputs": { "churn_rate": 0.02 } } ],
  "criteria": [ { "metric": "ending_mrr", "weight": 3 } ],
  "stress": { "variable": "churn_rate", "variation": 0.5, "steps": 5 }
} }

---

Deploy on Cloudflare Pages

npm install
npx wrangler login
npm run deploy      # esbuild -> site/_worker.js, then wrangler pages deploy

Runs fully free / private with no bindings or secrets (fails open on the free tier). For billing: bind a KV namespace AGENTSTACK_KV, set PRICE_STARTER / PRICE_PRO in [vars], and add Stripe secrets:

wrangler pages secret put STRIPE_SECRET_KEY    --project-name agentstack-mcp
wrangler pages secret put STRIPE_WEBHOOK_SECRET --project-name agentstack-mcp

Routes: /mcp (Streamable HTTP), /checkout, /success, /portal, /webhook, /metrics.

Auth & rate limiting

Enforced in worker-src/billing.mjs: identify() reads X-API-Key / Authorization: Bearer and falls back to per-IP free tier; consumeQuota() is a KV daily counter (the single gating point). One key / one quota spans the whole stack. Marked NOTE (auth) / NOTE (rate limiting) seams show where to swap in JWT/mTLS or a token-bucket. None of it touches the deterministic math.

Relationship to the standalone servers

AgentStack is additive, not a replacement. The three products keep running on their own for single-domain use:

• ScenarioSim — https://scenariosim-mcp.pages.dev/mcp
• DecisionMatrix — https://decisionmatrix-mcp.pages.dev/mcp
• PrecisionCalc — https://precisioncalc-mcp.pages.dev/mcp

License

MIT — see LICENSE.