Claude remembers your projects, your writing style, the context you've shared across conversations. It does not know what you spent last quarter. So when you ask it for budgeting advice, the answer comes back as a generic checklist: "track your spending, make a budget, automate your savings." Useful for a 22-year-old. Less useful when the question you actually have is "did I overspend on travel last quarter?"

The fix is not better prompting. It is to give Claude real access to your real bank data, so the answer comes back with numbers from your accounts. This post walks through the setup end-to-end: the protocol that makes it possible (MCP), the bank connection (Plaid), and the FinContext server that bridges the two. Setup is about ten minutes. (Prefer ChatGPT, Cursor, Cline, or Zed? The same server works there too — more on that below.)

What MCP changes

The Model Context Protocol (MCP) is a standard for letting AI clients — Claude Desktop, ChatGPT with custom connectors, Cursor, Cline, Zed — read data from external tools through a uniform interface. Before MCP, every "AI plus your data" integration was bespoke: a ChatGPT plugin here, a Claude tool there, a different config for every editor. MCP collapses that surface area: one server speaks the protocol, and every MCP-compatible client can use it.

For personal finance, the practical effect is that one MCP server can serve your bank context to whatever AI you actually use. You are not locked into a vendor. You are not maintaining six integrations.

What FinContext does

FinContext is an MCP server for personal finance. It exposes ten tools that let an AI client read and lightly organize your bank context:

• status — account and onboarding state
• link_bank — Plaid Link URL to connect a new account
• balances — accounts and current balances
• transactions — search and list transactions
• categories — taxonomy and usage stats
• sync — pull the latest from Plaid's cache
• update — relabel a merchant or category
• stat — read-only SQL surface for arbitrary analytics
• feedback — send a note to the FinContext team
• help — describe any tool in detail

Six of those are read-only (status, balances, transactions, categories, stat, help). Four are writes that don't touch your bank (link_bank, sync, update, feedback — local relabeling and sync triggers). None move money. None have a code path that could move money. We come back to that.

What about ChatGPT's built-in finance?

Some assistants have started shipping built-in finance features — ChatGPT among them. Claude hasn't. So if Claude is your main assistant, FinContext is how you give it your real numbers today.

And because FinContext speaks MCP, it isn't tied to one assistant. You link your accounts once and the same connection works in Claude, ChatGPT, Cursor, Cline, and Zed — your bank context follows you to whatever AI you're using that day, instead of being rebuilt (or locked) inside each app. A built-in feature reads from whatever the platform has wired up; FinContext reads your bank — any of 10,000+ US institutions through Plaid.

It's also standalone at $4.99/month: it doesn't depend on which AI subscription you happen to be paying for, and it's read-only by architecture with delete-anytime control. The bet is simple — your bank data should be something you bring to any AI, not a feature you rent inside one.

Setup, three steps

The flow below uses Claude (Desktop or claude.ai). ChatGPT, Cursor, Cline, and Zed use the same custom-connector pattern; on Claude Code (the CLI) it's a one-line claude mcp add command. See the developers page for per-client instructions.

Step 1 — Sign up at fincontext.ai

Create an account. The free trial is full access with no credit card; pricing is on the homepage.

Step 2 — Connect a bank through Plaid

In your FinContext dashboard, click "Link a bank." Plaid's hosted flow opens — same flow you have probably used before for Venmo, Robinhood, or Chime. Choose your bank, log in, pick which accounts to share. FinContext never sees your bank credentials; Plaid handles authentication and returns a scoped, read-only access token, which we encrypt and store.

Step 3 — Add FinContext to Claude as a custom connector

In Claude (Desktop or claude.ai), open Settings → Connectors → Add custom connector. Enter:

https://fincontext.ai/mcp

Claude redirects to FinContext to log in and grant access. There is no API token to paste — the OAuth flow handles authentication. After approval, Claude shows the connector as active. (Using Claude Code, the CLI? It's a one-line claude mcp add with your API token — see the developers page.)

That is the whole setup.

Try these prompts

The tools are documented; the prompts are how you actually use them. These five are the canonical prompts from our own testing — the questions we built the product to answer.

1. Monthly spending review. "How am I doing this month?" Claude calls stat for current-month spending by category, then stat for the 3-month historical average, then balances for the current snapshot. The answer comes back something like: month-to-date $1,019 across six categories, with rent not yet posted, and 3-month averages for comparison.

2. Subscription audit. "Am I wasting money on subscriptions?" Claude queries stat for recurring merchants — defined as ≥3 charges in 12 months with low amount variance — then sorts the result into SaaS subscriptions vs. fixed life costs (rent is recurring; rent is not Netflix). On our test account, this surfaces three or four real SaaS subscriptions and confirms nothing has been quietly auto-renewing into the void.

3. Affordability check. "Can I afford a $2,000 vacation next month?" balances for liquid totals, stat for 6-month average income and spending, then arithmetic: surplus per month, runway covered, whether the trip eats into savings. The answer is a number with a one-paragraph rationale, not a hedge.

4. Spending diagnosis. "Why does it feel like I'm spending more?" This one is interesting because the answer is often "you're not." stat compares current month by category against the 3-month baseline; if a category is over baseline, Claude pulls top merchants in that category. Sometimes the diagnosis is one $114 Shell charge. Sometimes it's that rent has not posted yet.

5. Net worth progress. "Am I making progress?" balances for current, stat for the historical balance series. The answer reports the trajectory — and explicitly flags reconstructed history (computed by walking transactions backward) as estimate, not audited snapshot. Real daily snapshots accrue from when you connect.

The bigger point: an AI assistant with FinContext is not a dashboard. You phrase the question in your own terms — your definition of "this month," your sense of "subscriptions" — and the AI translates that into the right tool calls.

What FinContext can read, what it cannot

The trust question deserves a direct answer.

Can read: account names and types, current balances, transaction history (merchant, amount, date, Plaid category), and any merchant or category overrides you have set in FinContext.

Cannot read: your bank password (Plaid handles authentication; we never see it), your conversations with Claude or ChatGPT, anything outside transactions and balances, or anything from a bank you have not explicitly connected.

Cannot do, by architecture, not policy: move money, place trades, initiate transfers, or take any action against your bank. There is no code path in the FinContext server that does any of those things. The Plaid scope we request is read-only; even if the server were compromised, the attacker has no privileged action available to them.

Storage: Plaid access tokens are encrypted at rest with Fernet (authenticated AES-128-CBC plus HMAC-SHA256); the key is stored in Google Cloud Secret Manager and injected at runtime — never in the database or the codebase. Everything travels TLS 1.3. Each user's data is isolated by Postgres row-level security — not by application-level filters that can have bugs. You can disconnect a bank, delete the account, or set a 30-day retention window at any time.

US-only today.

Limitations worth flagging

Honest about the rough edges:

• Sync latency. FinContext pulls from Plaid's cache. Transactions posted in the last hour or two may not show up until the next sync cycle. Run the sync tool to force-pull from cache; it does not force the bank to re-poll.
• Reconstructed history. Daily balance history before your account-link date is reconstructed from transaction flow, not pulled from the bank. Treat older points as estimates.
• AI hallucination. The data is real, but the AI can still misread or oversimplify it. For decisions that matter, confirm by asking Claude to show its work — it can repeat the underlying numbers from the tool calls.

Where to go next

If this worked for you, the same MCP endpoint works in ChatGPT, Cursor, Cline, and Zed too — same setup pattern, different add-connector flow. Link once, use your bank context anywhere. The developers page has per-client instructions.

If you want to dig into the architecture before connecting a bank, I wrote up the full threat model separately — Is it safe to connect my bank account to AI? — and the security page documents the data inventory.

Try it: fincontext.ai.