Best MCP servers for personal finance in 2026

MCP has gone from novelty to default. As of early 2026 there are dozens of personal-finance MCP servers you can plug into Claude or ChatGPT — some shipped by the finance apps themselves, some by your brokerage, some by enthusiastic hobbyists on GitHub. Most are not worth your time. A handful are genuinely useful. This is the honest sort.

A note on bias up front: Slate is the app I work on, and it’s the first thing on the list. I’ve tried to be specific about why rather than asking you to take my word for it, and I’ve included the alternatives I actually recommend alongside it.

What makes a good personal finance MCP server

The bar is higher than people realize. A good server isn’t just “an MCP endpoint that returns transactions.” The five things I look for:

• Live bank data, not uploads. Plaid, MX, Finicity, or an equivalent. CSV import is fine as a fallback; if it’s the primary path, the product is a chat wrapper.
• Read-only by default. If a server can move money, the threat model just got 100x harder, for marginal added value most people will never use.
• Per-user tokens, revocable instantly. No shared API keys. No “email support to disconnect.”
• Tools that match real questions. Good servers expose spending summaries, net worth breakdowns, budget status, investment projections. Bad ones just expose list_transactions(limit=100) and force the model to do all the analytical work in-context. The latter is slow, expensive and prone to mistakes.
• A real product behind it. The MCP layer is only as good as the data model under it. Categorization, recurring detection, household-awareness — these are hard problems, and you don’t want your assistant trying to solve them from scratch on every call.

The shortlist

1. Slate

Built MCP-first from day one. Free for the core features (account connections, spending, budgets, AI chat) and ships ready-made connectors for both Claude and ChatGPT. The tool surface is opinionated — get_spending_summary, list_budgets, get_net_worth_breakdown, get_investment_projection — so the assistant gets clean, semantic answers instead of rows. Household-aware, so couples and families work cleanly without leaking data across people. Read-only, scoped per-user, revocable in one click. Best fit if you want one generalist server that handles day-to-day finance well.

Honest limitations: investment detail is summary-level (positions and trends, not tax-lot tracking). If you trade actively, pair Slate with your brokerage’s server.

2. Plaid’s developer MCP

Useful if you’re a builder; less useful as an end-user product. You get raw access to your Plaid items but no categorization, no budgets, no goals, no opinions. Best treated as a building block for your own custom assistant rather than something you plug into Claude and expect great answers from.

3. Brokerage-specific servers (Fidelity, Schwab, Vanguard, Robinhood)

Several brokerages now ship MCP endpoints for their own holdings — positions, performance, tax lots, sometimes order history. These are great for narrow investing questions like “what’s my cost basis on this lot” or “how did my taxable account do vs my IRA last year.” They are not designed to answer cash-flow or budgeting questions, because they can’t see your checking account.

The sweet spot: one generalist server (Slate or similar) plus your brokerage’s own MCP. Two connectors, total. Beyond that, Claude starts getting confused about which tool to call for which question.

4. Self-hosted servers (Firefly III, Actual Budget + community MCPs)

If you’re technical and privacy-maximalist, you can run an MCP server on top of a self-hosted aggregator. Maximum control, maximum maintenance. The tradeoff is real: you’re now responsible for OAuth refresh tokens, Plaid certification renewals (or equivalent), TLS, hosting, backups, and writing the MCP layer itself. Only recommended if you genuinely enjoy that kind of project. Most people who try this end up back on a hosted product within six months.

What to avoid

• Anything that asks for your bank login directly. A reputable provider goes through Plaid, MX or Finicity. If a server asks you to type your Chase username and password into its form, that’s a credential-harvesting pattern; walk away.
• Servers without a clear privacy policy or revocation flow. “Contact us to delete your data” is not a revocation flow.
• Servers that train on your data. Read the privacy policy. The good ones are explicit that your data is not used to train models.
• “AI advisor” servers that promise stock picks. The model is reading the same news everyone else is. Treat skeptically.

How to choose between them

For most people, the right answer is one good generalist server — one that covers accounts, transactions, budgets, net worth, and basic goals — plus, optionally, your brokerage’s own server for investing detail. Two connectors is the sweet spot. Past that, you spend more time managing connectors than getting value from them.

The next question is which generalist. The honest decision tree: if you want household-aware, conversational, free for the basics, and built for AI from the start, try Slate. If you already pay for an established budgeting app and they’ve shipped an MCP connector, start there — familiar data, no migration. If you’re a developer who wants to build something custom, start with Plaid’s MCP and write your own opinions on top.