How to give Claude real-time access to your business finances with MCP

What is MCP?

MCP (Model Context Protocol) is an open standard that lets AI assistants call external tools. Instead of pasting data into a prompt, you give the model direct access to an API. It decides when to call which tool based on your question. FlowCheck ships an MCP server that exposes your financial data as tools Claude can call in real time.

What you need

• A FlowCheck API key (get one here, free trial included)
• At least one data source connected (Stripe, Shopify, or a bank account)
• Claude Code (CLI) or Claude Desktop, depending on which setup you prefer

Option 1: Claude Code (stdio transport)

This is the fastest setup. One command adds FlowCheck as an MCP server that runs locally via stdio:

claude mcp add flowcheck \
  --transport stdio \
  --env FLOWCHECK_API_KEY=fc_live_your_key \
  -- npx -y @flowcheck/mcp-server

That is it. Claude Code now has access to your FlowCheck data. Start a conversation and ask a question:

> How much revenue did we make this month?

Claude will call the FlowCheck cashflow tool and respond:

"Based on your FlowCheck data, your total inflow for March 2026
is $12,847.00 across 23 Stripe payouts and 8 Shopify payouts.
After $4,230.00 in outflows, your net cash flow is $8,617.00."

Option 2: Claude Desktop (JSON config)

If you use Claude Desktop, add the server to your MCP configuration file. On macOS this is ~/Library/Application Support/Claude/claude_desktop_config.json. On Windows, it is in %APPDATA%\Claude\claude_desktop_config.json:

{
  "mcpServers": {
    "flowcheck": {
      "command": "npx",
      "args": ["-y", "@flowcheck/mcp-server"],
      "env": {
        "FLOWCHECK_API_KEY": "fc_live_your_key"
      }
    }
  }
}

Restart Claude Desktop. You will see FlowCheck listed in the MCP tools panel.

Option 3: Remote MCP endpoint

For clients that support remote MCP servers (Glama, Smithery, custom agents), point to the hosted endpoint:

https://developer.usepopup.com/api/mcp

Pass your API key as a Bearer token. The remote server exposes the same tools as the stdio version but runs on FlowCheck infrastructure instead of your local machine.

Available tools

The MCP server exposes these tools to the AI model. Claude picks the right one based on your question:

Tool	What it does
get_balance	Stripe available/pending + bank balances
get_cashflow	Daily inflow/outflow over 7, 30, or 90 days
list_payouts	Stripe + Shopify payouts with reconciliation status
get_reconciliation	30-day match rate and discrepancy count
list_discrepancies	Missing deposits, amount mismatches, timing alerts
get_position	7-day financial summary (balances + recent payouts + cash flow)
get_alerts	Active issues in agent-friendly format

Example prompts

Once connected, you can ask Claude natural questions about your finances. Here are some that work well:

• "What is our current cash position?"
  Claude calls get_balance and returns Stripe available, Stripe pending, and bank balance with a total.

• "Are there any unmatched payouts this month?"
  Claude calls get_reconciliation and lists any payouts that have not been matched to a bank deposit, with amounts and dates.

• "Compare last week's revenue to the week before."
  Claude calls get_cashflow with a 30-day window and computes the week-over-week change from the daily breakdown.

• "Summarize our financial health for the board meeting."
  Claude calls get_position and formats the response as a brief executive summary with revenue, expenses, net, and any open issues.

Sandbox mode

If you want to test without touching production data, use a sandbox key (prefixed fc_test_). Sandbox calls return realistic test data and do not consume credits. This is useful for building agents or testing prompts before going live.

claude mcp add flowcheck-sandbox \
  --transport stdio \
  --env FLOWCHECK_API_KEY=fc_test_your_key \
  -- npx -y @flowcheck/mcp-server

Get started

Get an API key, connect at least one data source from the dashboard, and run the claude mcp add command above. The whole setup takes under five minutes. Once connected, ask Claude anything about your business finances and it will pull the answer from your live data.

Comments

[Hamza KONTE]

Clean implementation. The MCP pattern for financial data is particularly well-suited here because you're solving the "freshness problem" — financial data goes stale in minutes and is painful to paste manually, so the tool-call approach is genuinely better than context stuffing.

One thing worth noting for anyone building on top of this: the prompt you use to query Claude matters a lot for financial questions. Asking "how are we doing financially?" will get a different (often vaguer) answer than asking "compare our net cash flow this week vs. last week and flag any single outflow over $5k." The tool handles data retrieval, but the question structure determines what the tool gets called with and how the result is framed.

For anyone building agents on top of MCP servers like this — the pattern of stacking a prompt-structuring MCP with a domain-specific data MCP works well. One handles "what should I ask and how" and the other handles "fetching the answer." Cleaner separation than trying to put financial reasoning into the tool descriptions themselves.

[Hamza KONTE]

This is a great use case for MCP — giving Claude grounded, real-time context rather than relying on stale training data for financial decisions.

One thing I'd add: the quality of what Claude does with that financial data depends heavily on how you instruct it. Vague prompts like "analyze my finances" produce vague summaries. But a structured prompt with clear objective (identify cash flow anomalies), output format (table with date/amount/category/flag), and constraints (only flag items > $500) produces actionable output.

I built a free tool called flompt (flompt.dev) specifically for this kind of prompt structuring. It breaks prompts into semantic blocks and compiles them into Claude-optimized XML. Pairs really well with MCP setups like yours — the data pipeline is great, but the prompt layer is where precision comes from.