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 agent 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) work with structured data from third-party tools through a shared protocol. Ledg uses a remote HTTPS endpoint. 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. With your approval it can also hand a receipt back the other way.

This guide explains what MCP is, why it matters when you run a Canadian business on your own, and how to connect it and check that it is reading the right business and period.

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.

Ledg exposes tools including these (the client supplies the business ID and required date parameters):

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

Tool results give the model recorded data to work from, but do not guarantee a correct answer. Check the selected business, dates, filters and source entries. Missing records and misclassified transactions can still produce an incomplete total, and the model can misinterpret a result.

Why this matters when you run the business yourself

Most Canadian solo 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?" means logging into the bank, checking three accounts, and summing them in your head. "How much did I make last quarter?" means logging into the accounting tool, running a report, and squinting at a 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.

Six things worth asking

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

1. "How much GST/HST did my recorded sales collect this quarter?"

The AI can call get_sales_tax_collected for the selected business and dates. This is recorded tax collected, not a complete GST/HST return. The expense-total tool does not supply a verified ITC calculation. Reconcile eligible ITCs, adjustments and your reporting method before filing; Ledg does not submit the return.

2. "Any uncategorized transactions this week?"

The AI calls find_uncategorized_transactions and lists the items needing your attention. The tool returns up to 100 recent matching entries. Review them in Ledg. Reminders require a separate scheduling feature in your AI client; this Ledg tool does not create them.

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

The AI first checks list_categories, then calls get_expenses with the exact category and date range. A recorded expense total is not the final tax deduction: business meals generally face a 50% limit, with exceptions explained in the meals guide.

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

The AI discovers account IDs with list_accounts, reads their recorded balances with get_account_balance, and filters list_deadlines. Ask it to show the as-of date and distinguish a bookkeeping balance from a live bank balance. Due dates depend on the business profile you entered; a deadline does not establish the amount owing.

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.

6. "Here is a photo of a cash receipt. Record it."

The purchase is drafted into Stage with its tax breakdown and line items. It stays out of your statements and tax totals until you confirm it in Ledg.

How to set it up

Claude.ai or Claude Desktop

1. Sign up for Ledg, switch to Solo (the AI agent connection is a Solo feature, see pricing), and import a few transactions so there's data to read.
2. Add Ledg to Claude. The link opens Claude's Add custom connector dialog with Ledg's name and URL filled in; select Add, then Connect. By hand, open Customize, then Connectors, and use the plus button to add a custom connector. A Team or Enterprise owner may need to add it for the organization first. Claude Free allows one custom connector.
3. If you added it by hand, paste the MCP URL: https://ledg.ca/api/mcp
4. Approve the OAuth grant in the popup. Claude opens Ledg, you select Approve and continue, and you land back in Claude. Done.
5. Enable the connector in the conversation and ask one of the questions above. Check that Claude actually used the relevant Ledg tool. See Claude's current connector instructions.

ChatGPT

OpenAI's current MCP connection instructions use Developer mode. It needs ChatGPT Plus, Pro, Business, Enterprise or Edu, set up on chatgpt.com in a browser; ChatGPT Free and Go do not have it.

1. In ChatGPT settings, open Security and login and turn on Developer mode. The plus button in Plugins only creates an app once this is on.
2. Open ChatGPT Plugins, select the plus button (Add plugin), then Create app, and name it Ledg.
3. Paste https://ledg.ca/api/mcp as the MCP server URL and keep OAuth for authentication. Tick the box under the risk warning and select Create.
4. Complete Ledg's authorization, reviewing the requested permissions.
5. In every message, pick Ledg from the plus button in the message box, or start the message with @Ledg. ChatGPT uses Ledg only in a message where you pick it. Ask a read-only question first to verify the business and date range.

If these controls are unavailable, check your workspace administrator and OpenAI's linked instructions. The Ledg URL alone does not make tools available on every account.

Cursor and other power-user clients

For non-OAuth clients, generate a Personal Access Token (PAT) from Ledg's Settings, then AI agents, then Advanced. The token starts with ledg_mcp_ and goes into your client config as a Bearer token. Access tokens are read-only; recording a receipt needs the OAuth connection.

What about privacy?

The connection reads your books, and with your approval it can record a purchase or add a receipt's details. Anything it records lands in Stage and touches no statement or tax total until you confirm it in Ledg. It cannot confirm an entry, change one you already confirmed, or move money. Revoke it any time in Settings. The current tool list is in the setup guide.

The AI provider receives data returned by the tools you authorize. Its retention and training rules depend on the product, plan and your settings. Review those controls before connecting financial records. Revoking access prevents future access through the connection; it does not erase copies already received by the provider.

Will this kill the accountant?

No. The MCP layer answers "what's the state of my books right now?" That's the owner's question, asked whenever it comes up. 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. Compatibility also depends on the client's supported transport, authentication and protocol features. 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 work with 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 uses OAuth in supported clients, or a read-only PAT in clients that accept a Bearer token.
• Reads your books and can record receipts into Stage. Nothing counts until you confirm it in Ledg.
• Doesn't replace the accountant; replaces the dashboard for quick questions.

Connect your AI agent