Finance Manager MCP
Enables self-hosted personal finance management for MCP clients by connecting read-only bank accounts via SimpleFIN, syncing transactions into a local SQLite database, and exposing tools to query balances, spending, budgets, cash flow, and net worth.
README
Finance Manager MCP
Self-hosted personal finance for MCP clients. Your banks stay read-only (SimpleFIN), your data lives in a local SQLite file you own, and an LLM can only query deterministic totals — it never writes the ledger.
SimpleFIN (read-only)
→ raw_syncs (append-only JSON)
→ truth tables (accounts, transactions, rules, budgets)
→ warm-cache aggregates
→ atomic functions / MCP tools (stdio + gRPC)
Why this shape
- Idempotent ingest — transactions are keyed on the bank-assigned SimpleFIN id. Re-syncing the same dump yields zero net-new rows; identical-looking new charges are never collapsed.
- Deterministic categorization — first-match-wins regex rules. No model in the write path.
- Warm cache — monthly spending, cash flow, net-worth snapshots, and recurring streams persist in SQLite so restarts are instantly warm. Cache is derived;
rebuild-cacheregenerates it from truth tables. - Profile UUID namespacing — every write tool requires a
profile_id. Multiple profiles never leak into each other. - Dual transport — stdio for Claude Desktop / Cursor, gRPC via the sibling GRPC-MCP package.
Setup
Python 3.11+.
python -m venv .venv
# Windows: .venv\Scripts\activate
# Unix: source .venv/bin/activate
pip install -e ".[dev]"
# optional gRPC transport (sibling ../GRPC-MCP)
pip install -e "../GRPC-MCP/python"
pip install -e ".[grpc]"
cp .env.example .env
# paste a Fernet key into FINANCE_ENCRYPTION_KEY
python -c "from cryptography.fernet import Fernet; print(Fernet.generate_key().decode())"
finance-mcp init # creates data/finance.db + a default profile
finance-mcp profile list # copy the profile UUID
Point it at SimpleFIN
From an MCP client, ask to connect a bank. The agent should:
- Call
begin_simplefin_connect(or use promptconnect_simplefin). - Show you bridge.simplefin.org — you sign in, connect banks (read-only), copy the one-time setup token.
- Call
claim_connectionwith that token. The access URL is encrypted at rest. - Optionally
sync_nowon the sameprofile_id.
CLI still works if you prefer a terminal:
finance-mcp claim-token --token "<setup-token>" --name "My Banks" --profile <PROFILE_UUID>
finance-mcp sync --profile <PROFILE_UUID>
sync is safe to run from Task Scheduler / cron. Overlapping windows do not double-count.
Serve
finance-mcp serve-stdio # Claude Desktop / Cursor
finance-mcp serve-grpc --auth bearer # sibling GRPC-MCP transport
finance-mcp token new --name cursor --profile <PROFILE_UUID>
The bearer token is printed once. gRPC stores only the SHA-256 hash.
Atomic functions
Everything in finance_mcp.api is a plain function. MCP tools are thin wrappers. Import the library without running a server:
from finance_mcp.config import Settings
from finance_mcp.core.db import Database
from finance_mcp import api
db = Database.from_settings(Settings.from_env())
print(api.net_worth(db, profile_id))
Read
list_accounts, get_balances, net_worth, net_worth_history, search_transactions, recent_transactions, spending_by_category, cash_flow, recurring_streams, upcoming_payments, budget_status, list_rules, sync_status, profile_info, list_profiles, begin_simplefin_connect
Write (always require profile_id)
claim_connection, create_rule, update_rule, delete_rule, apply_rules, create_budget, update_budget, delete_budget, sync_now
Tests
The suite is hermetic — synthetic fixtures only, no credentials.
pytest -q
Security
- SimpleFIN access URLs are encrypted with
FINANCE_ENCRYPTION_KEY(Fernet). Back the key up. - gRPC bearer tokens are stored as hashes.
- Read tools open SQLite with
mode=ro. Write tools are the only mutation path and are profile-scoped. data/and.envare gitignored.
License
Use as you like for personal self-hosted finance.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。