How to Connect Your Bookkeeping to Claude or ChatGPT (MCP Guide for Canadians)

For most of accounting software's history, the answer to "how much did I spend on meals last quarter?" required either opening a dashboard, running a report, or firing up a spreadsheet. In 2026 there's a third option: ask the AI client you already have open.

This works because of the Model Context Protocol (MCP), an open standard that lets AI clients (Claude.ai, Claude Desktop, ChatGPT, Cursor) read structured data from third-party tools through a normalized HTTP layer. If your bookkeeping tool ships an MCP server, your AI client can ask it for sales numbers, expense breakdowns, GST positions, account balances, and upcoming deadlines, and quote them back to you with real numbers.

This guide explains what MCP is, why it matters for solo Canadian operators specifically, and how to wire it up in three minutes.

What MCP actually is

Think of MCP as a USB-C port for AI clients. The AI client (Claude, ChatGPT) is the laptop. The bookkeeping tool (or notes app, calendar, CRM, anything) is the peripheral. MCP is the cable + plug spec that lets them talk without each pair needing a custom integration.

Practically, an MCP server exposes a list of named tools that the AI client can call. Each tool has a description, input parameters, and a return value. When you ask Claude a question, Claude looks at the available tools, picks the one it thinks is relevant, calls it, and uses the response to answer you.

A bookkeeping MCP server typically exposes tools like:

• get_sales(period) — total revenue between two dates
• get_expenses(period, category?) — total expenses in a period, optionally filtered by category
• get_account_balance(accountId, asOf) — balance of a specific account on a date
• find_uncategorized_transactions() — entries waiting for review
• list_deadlines() — upcoming tax filing deadlines
• ...and so on

The AI client doesn't make up numbers; it calls the right tool and quotes the result back. No hallucinated GST collected, no "I think your meals expense was around $X" guessing.

Why this matters for solo operators

Most 1-person Canadian businesses don't have time for "operations". You don't open the dashboard daily. You don't have a controller. You wear all the hats, which means the accounting hat is the last one on, usually right before the GST deadline or year-end.

The friction this creates: every question about your business requires a context switch. "How much cash do I have?" → log into bank → check three accounts → mentally sum. "How much did I make last quarter?" → log into accounting tool → run report → squint at chart. These are tiny tasks but they cost focus, and they're why most owners have a vague rather than precise sense of where their business stands month to month.

MCP collapses that. You're already chatting with Claude or ChatGPT for other reasons (drafting an email, debugging code, reviewing a contract). Asking "what's my cash position right now?" inline is zero context switch.

Five questions worth asking

These are real prompts you can use once your bookkeeping is connected via MCP.

1. "What's my GST position for the current quarter?"

The AI calls get_sales and get_sales_tax_collected for the current quarter, plus get_expenses and the corresponding ITC math. Returns the GST34-2 line numbers (line 101 for revenue, 105 for tax collected, 106 for ITCs claimed, 108 net). When the quarterly deadline arrives, you paste these into CRA My Business Account.

2. "Any uncategorized transactions this week?"

The AI calls find_uncategorized_transactions and lists the items needing your attention. Usually three or four bank charges or weird payment processor entries. Categorize them in two minutes via the dashboard, or just tell the AI what each one is and have it remind you next week if you want to batch.

3. "How much have I spent on Meals & Entertainment year-to-date?"

The AI calls get_expenses(category="Meals & Entertainment") for YTD. Quotes the number, plus the reminder that meals are 50% deductible against business income.

4. "What's my cash position and what deadlines are coming up in the next 30 days?"

The AI calls get_account_balance for each of your accounts and list_deadlines for the upcoming month. Single answer: "$42,310 across two chequing accounts. Q1 GST return due in 18 days (~$2,800 owing); BC Annual Report fee due in 25 days ($43)."

5. "Did I make more this October than last October?"

The AI calls get_sales twice with two date ranges and returns the comparison plus the percentage change. Useful for the "am I growing?" gut check that doesn't merit a chart.

How to set it up (three minutes)

Claude.ai or Claude Desktop

1. Sign up for ledg and import a few transactions so there's data to read.
2. In Claude, open Settings → Connectors → Add custom connector.
3. Paste the MCP URL: https://ledg.ca/api/mcp
4. Approve the OAuth grant in the popup. Claude redirects to ledg, you click "Allow", you bounce back. Done.
5. In any conversation, ask one of the five questions above. Claude picks the right tool automatically.

ChatGPT (Plus / Team / Enterprise)

1. In ChatGPT, open Settings → Connectors → Add MCP server.
2. Paste https://ledg.ca/api/mcp.
3. Approve the OAuth flow. Done.
4. Tools show up in the conversation immediately.

Cursor and other power-user clients

For non-OAuth clients, generate a Personal Access Token (PAT) from ledg's Settings → Connections → Advanced. The token starts with ledg_mcp_ and goes into your client config as a Bearer token.

What about privacy?

The MCP server is read-only by design. The AI client cannot post transactions, change categories, or move money. The 12 tools currently shipped (list_corporations, list_accounts, list_categories, get_sales, get_sales_tax_collected, get_expenses, get_account_balance, find_uncategorized_transactions, find_purchases_without_pst, list_deadlines, list_entry_attachments, list_minute_book) all read state and return it.

The AI client receives the data you ask for and can use it to answer your prompt. Whether that data is then logged or trained on depends on the AI client's policy (Claude doesn't train on your conversations by default; ChatGPT depends on the workspace setting). Read your AI client's data handling policy to be sure.

Will this kill the accountant?

No. The MCP layer answers "what's the state of my books right now?" That's the operator's question, asked daily. The accountant's question is "given the year's books, what does my T1 + T2125 (sole prop) or T2 + Schedules 1/8/100/125 (corp) look like?" That's a different kind of work and requires the accountant's judgment.

What MCP changes is who reads the books. Today it's the owner who logs into the dashboard. Tomorrow, it might be the accountant via their own AI client during a quick review, asking "any unusual expense categories this quarter?" instead of scrolling through 800 transactions.

What about other AI tools?

The Model Context Protocol is open. Any AI tool that implements the spec can connect to any MCP server. The current set is Claude (Anthropic), ChatGPT (OpenAI), Cursor, and a growing list of niche clients. Expect this list to broaden through 2026.

TL;DR

• MCP is an open standard that lets AI clients read data from third-party tools.
• Connecting your books means questions about your business get answered in chat without opening a dashboard.
• Real questions, real numbers: GST position, expense breakdowns, cash, deadlines, year-over-year comparisons.
• Setup in three minutes via OAuth (Claude / ChatGPT) or a PAT (Cursor / scripts).
• Read-only. The AI cannot change anything in your books.
• Doesn't replace the accountant; replaces the dashboard for the daily-question case.

Connect ledg to Claude or ChatGPT →