MCP for Claude
Ask Claude about your money.
Pare speaks the Model Context Protocol, exposing your finance data as 20tools an MCP client can call. On the hosted service that's a one-URL connector you add to claude.ai (web, desktop, or mobile); self-host ships a fully local server for Claude Code and Claude Desktop. Connect it once and you can ask Claude about your spending, budgets, and subscriptions in plain language, with the answers grounded in your actual transactions.
How it works
MCP is an open standard for giving an AI assistant access to tools and data. Claude calls a tool like spending_summary, Pare reads your database, and hands back the numbers. Claude then explains them in conversation.
Hosted: add Pare as a custom connector in claude.ai — one URL, an OAuth sign-in with explicit consent, no terminal. Every tool call runs against your own isolated database and nothing else; you can revoke access at any time.
Self-host: the server runs on your own machine over stdio and makes no network calls — it reads and writes only the local database, bypassing the web app entirely.
What you can ask
- > How much did I spend on restaurants last month?
- > What subscriptions am I paying for?
- > Set a $400 restaurant budget.
- > I spent $40 cash at the market.
- > Which categories are pacing over budget this month?
- > Show me my biggest one-off purchases this year.
- > Am I on track to finish the month with a surplus?
The tools
Ten read tools answer questions; eight write tools let Claude set goals, tune categorization rules, and re-tag transactions — so you can keep your data organized by asking, not clicking.
READ · 11
- spending_summary — monthly totals, category breakdown, top merchants
- list_transactions — filter by category, source, flow, date range, search
- category_breakdown — spend per category, optionally one month
- income_summary — income by type + income-vs-spend series
- cashflow — net cashflow per month + period surplus
- baseline — discretionary baseline with large one-offs removed
- subscriptions — detected recurring charges + double-bill flags
- goals_status — goal progress for the latest data month
- insights — auto tips: over-budget, MoM moves, one-offs
- list_categories — categories in use + keyword rules
- list_statements — uploaded/synced statements with ids + balances
WRITE · 9
- set_goal — set a category's monthly limit
- delete_goal — remove a category's monthly limit
- add_category_rule — add a keyword→category rule
- delete_category_rule — remove a keyword→category rule
- recategorize_all — re-apply rules to all transactions
- tag_transaction — override one transaction's category
- add_manual_transaction — record a cash / off-statement purchase
- delete_manual_transaction — delete a manually recorded transaction
- delete_statement — delete a statement + all its transactions
Privacy
The tools touch only your own database — local when you self-host, your isolated per-account store on hosted — and upload nothing on their own. Be aware, though, that whatever an AI client reads through these tools is sent to that client's model provider as conversation context — so connect only clients you trust with your financial data.
Set-up guides by client
Step-by-step instructions for each client, with the exact config file and the mistakes that break setups:
Setting it up
Hosted: paste one URL into claude.ai's Settings → Connectors. Self-host: add a small block of JSON to your Claude Code or Claude Desktop config. Either way, the Connect page shows the exact, copy-paste-ready setup for your account once you're signed in — including a one-line smoke test for the local server.