I wanted to ask my AI assistant one boring question: "where did my money go last month?" Not a budget template, not a categorization chore. Just an answer, with the transactions behind it.
So we built Sprout, a remote MCP server (plus iPhone, Mac and web apps) that connects US and Canadian bank and card accounts through Plaid, read-only, and exposes them to Claude, ChatGPT, Cursor or any MCP client. Here is what mattered once real money data was involved.
1. Read-only is a product decision, not just a scope
Every tool except one is read-only, and the one exception (update_settings) only saves a default reporting period. There is no tool that can move money, pay a bill or cancel a subscription, so there is nothing for a prompt injection inside a transaction description to trigger. We also treat transaction text as data, never instructions: merchant names come from banks and can contain anything.
2. Let the server do the math
LLMs are bad at summing 400 transactions. Each tool returns computed figures, not raw rows to add up:
-
get_cash_flow– income vs spending for a period -
get_spending_breakdown– by category or merchant -
get_period_totals/compare_periods– trends and month-over-month -
search_transactions– a specific charge or merchant -
get_recurring_charges– subscriptions and bills -
find_unusual_transactions– duplicates and outliers -
list_accounts,get_data_health– what is connected and how complete it is
Two rules saved us from wrong answers: transfers between your own accounts are excluded from spending, and USD and CAD are never summed together.
3. "Missing data" is not "zero spending"
If an account only synced 20 days, "you spent $0 on groceries in the first 10 days" is a lie. get_data_health tells the model what is connected and how far back data goes, and every answer carries the date range it used, so the user can check it against their bank.
4. Tool descriptions are routing logic
Most bad answers we saw were the model picking the wrong tool. Naming sibling tools inside each description ("for a single merchant use search_transactions instead") fixed more issues than any prompt change.
5. OAuth in 2026: support CIMD and dynamic client registration
Claude and ChatGPT use Client ID Metadata Documents. Plenty of other clients (mcp-remote, gateways, some IDE agents) still only do dynamic client registration. If your authorization server only advertises one, half the ecosystem silently can't sign in. Turn both on.
6. Distribution: publish to the official registry first
We published group.mindcrafters/sprout to the official MCP Registry with a DNS-verified namespace. Several directories ingest from it, so one publish seeded multiple listings.
Try it
- MCP endpoint:
https://api-sprout.mindcrafters.group/mcp(OAuth, no API key) - Plugin manifests (MIT): github.com/MindCraftersGroup/sprout-claude-plugin
- iPhone/iPad/Mac: App Store
Ask it "compare my spending in October with September" and tell me what breaks. I'm especially curious which MCP clients still fail the OAuth handshake.
Top comments (0)