Personal Finance MCP Server

一個可整合至 Claude 系統的 MCP 伺服器，協助您在 Claude 生態系統中更精準地計算個人財務。

文件

💰 Personal Finance MCP

Deterministic personal-finance toolkit exposed over the Model Context Protocol — 77 calculators, a meta-advisor, and live market data, with a polished web UI. Grounded in established financial mathematics.

Live demo: https://sarveshtalele-personal-finance-mcp.hf.space Connector URL: https://sarveshtalele-personal-finance-mcp.hf.space/mcp

Demo

▶️ Watch the 2-minute demo — plain-language question → chained tools → a prioritized plan.

Public demo note: the hosted Space is a shared, best-effort instance (rate-limited, may cold-start after idle). For heavy or private use, run it locally or self-host (see docs/HOW_IT_WORKS.md).

Overview

Most finance "assistants" guess at numbers. This one doesn't. It ships 77 deterministic calculators — same inputs, same answer, every time — and lets an LLM route a plain-language question to the right tools. Describe your situation ("I'm 30, earn ₹1L/month, want to retire at 60") and the create_financial_plan orchestrator chains the relevant calculators into a single prioritised plan.

It runs three ways from one codebase:

As an MCP server — connect it to Claude Desktop, Claude Code, Cursor, or any MCP client.
As a website — a Next.js UI with a live calculator, a market dashboard, and a tool catalog.
As a hosted connector — deployed to a Hugging Face Docker Space; one URL does all three.

Highlights

🔢 Deterministic — pure math, no model inference for the numbers.
🤖 Story → tools — the model maps intent to tools; users never name them.
🇮🇳 Theory-grounded — TVM, debt, PPF/SSY/NSC/EPF, bonds, derivatives, MPT, and more.
🛰️ Live market data — mutual-fund NAVs (AMFI), FX (ECB), equity quotes (Yahoo) — no API keys.
🔒 Hardened — stateless, rate-limited, input-bounded APIs with security headers/CSP.

Tool catalog — 77 tools, 13 categories

Category	Tools	Examples
Time Value of Money	10	future/present value, annuity, perpetuity, EAR, real return
Portfolio Analytics	11	CAPM, Sharpe, Sortino, Treynor, alpha, allocation, rebalancing
Financial Planning	9	net worth, ratios, emergency fund, retirement, education, insurance
Small Savings (India)	9	PPF, SSY, NSC, KVP, SCSS, RD, FD, EPF
Mutual Funds	7	SIP, SWP, lumpsum-vs-SIP, CAGR, NAV, expense-ratio impact
Debt & Loans	6	EMI, amortization, prepayment, consolidation, invest-vs-prepay
Fixed Income	6	bond price, YTM, current yield, duration, convexity, zero-coupon
Derivatives	5	futures fair value, option payoff, put-call parity, Black-Scholes, beta hedge
Equity Valuation	5	DDM, two-stage DDM, P/E, DCF, dividend yield
Live Market Data	4	MF search, live NAV, FX rate, stock/index quote
Cash Flow & Budgeting	3	household cash flow, debt-to-income, contingency fund
Risk Profiling	1	suitability score → suggested equity/debt split
Advisor	1	`create_financial_plan` — the story → plan orchestrator

Browse them all (with live descriptions) at /tools.

Quick start

Use the hosted connector (no install)

Claude Desktop — Settings → Connectors → Add custom connector → paste:

https://sarveshtalele-personal-finance-mcp.hf.space/mcp

Claude Code

claude mcp add --transport http personal-finance https://sarveshtalele-personal-finance-mcp.hf.space/mcp

Cursor / VS Code — add to mcp.json:

{
  "mcpServers": {
    "personal-finance": {
      "url": "https://sarveshtalele-personal-finance-mcp.hf.space/mcp",
      "transport": "http"
    }
  }
}

Install from PyPI (stdio server)

pip install personal-finance-mcp     # or: uvx personal-finance-mcp

Then point Claude Desktop at it:

{
  "mcpServers": {
    "personal-finance": { "command": "uvx", "args": ["personal-finance-mcp"] }
  }
}

Run locally from source

git clone https://github.com/sarveshtalele/personal-finance-mcp.git
cd personal-finance-mcp
pip install -e .

# Option A — classic stdio MCP server (offline, no web)
python -m src

# Option B — unified server: website + /mcp connector + /api  (http://localhost:7860)
cd web && npm install && npm run build && cd ..
python -m src.web

For stdio, point Claude Desktop at the local process:

{
  "mcpServers": {
    "personal-finance": { "command": "python", "args": ["-m", "src"] }
  }
}

The website

python -m src.web serves everything on one port:

Path	What
`/`	Next.js site — home, tool catalog, live calculator, market dashboard, setup guide
`/mcp`	MCP server over streamable-HTTP — the connector URL
`/api/*`	JSON endpoints (tool catalog, calculators, live market data)

Architecture

src/
├── server.py            # FastMCP server — registers all tool modules
├── __main__.py          # `python -m src`  (stdio transport)
├── tools/               # pure math fns + per-module register(mcp)
│   ├── tvm.py  debt.py  planning.py  bonds.py  stocks.py  mutual_funds.py
│   ├── portfolio.py  derivatives.py  india_savings.py  cashflow.py
│   ├── risk_profile.py  advisor.py   # advisor = story → plan orchestrator
│   └── marketdata.py    # live AMFI / Frankfurter / Yahoo (keyless)
├── models/              # Pydantic schemas + enums
├── utils/               # output formatters
└── web/                 # unified Starlette server (MCP + /api + static site)
    ├── server.py        # routes, security middleware, calculator registry
    └── __main__.py      # `python -m src.web`  (uvicorn, port 7860)

web/                     # Next.js front-end (static export → web/out)
└── app/                 # home, tools, calculator, dashboard, connect

Each tool file keeps deterministic pure functions separate from the thin @mcp.tool wrappers, so the same functions power the MCP server, the web calculators, and the tests.

Security

Stateless — no database, no sessions; every call is independent and reproducible.
Hardened API — per-IP rate limiting, request-body cap, and input validation that bounds loop-driving parameters (years/months/age) to prevent denial-of-service.
Security headers — CSP (with frame-ancestors for the Hugging Face embed), X-Content-Type-Options, Referrer-Policy, Permissions-Policy; CORS limited to GET/POST without credentials. The header middleware is implemented at the ASGI layer so it never buffers the streaming /mcp (SSE) responses.
No secrets in the app — live-data sources are public and keyless.

See SECURITY.md to report a vulnerability.

Development

pip install -e ".[dev]"
pytest -q                       # 119 tests
ruff check .                    # lint
python -m src.web               # run the full stack locally

Deployment

Deployed as a Hugging Face Docker Space, auto-synced from GitHub on every push to main (see .github/workflows/hf-sync.yml). Full instructions — local, Docker, and Hugging Face — are in docs/deployment.md.

Documentation

docs/HOW_IT_WORKS.md — concepts (MCP, transports, semantic routing), full architecture with diagrams, end-to-end request flows, the security model, hosting it yourself / on a portfolio site, and a production-grade roadmap.
docs/deployment.md — local, Docker, and Hugging Face deployment.
docs/Architecture.md · docs/testing.md · docs/setup.md

Contributing

Contributions are welcome — see CONTRIBUTING.md and the Code of Conduct. Good first issues: add a calculator (a pure function + a register wrapper + a test), improve descriptions for better tool routing, or extend the web UI.

Disclaimer

Educational tool for illustrating standard financial formulas. Not investment advice. Figures are illustrative; verify before making financial decisions.

License

MIT — free to use, modify, and distribute.