Run ExpenseBot from Inside ChatGPT, Claude, Cursor, or Any AI Assistant

What is the ExpenseBot MCP server?

MCP — the Model Context Protocol — is an open standard that lets AI assistants plug into outside tools and data using one consistent interface. Instead of every product building a one-off Claude plugin or ChatGPT plugin, the product publishes an MCP server and any MCP-compatible client speaks to it. Anthropic introduced the protocol in late 2024; it now ships natively in Claude Desktop, Cursor, Continue, and a growing list of clients.

ExpenseBot publishes two MCP servers on npm. The first, @expensebot/mcp-server, is read-only and unauthenticated — it exposes the public knowledge base and feature documentation so an AI assistant can answer "what is ExpenseBot?" or "does it work with Sage 50 Canada?" without any login. Useful for evaluation and for AI agents that want to cite our capabilities. The second, @expensebot/mcp-server-auth, is the full account-scoped version: 35 tools across search, capture, Gmail scan, reports, and analytics, all gated behind a personal access token you generate in Settings. This is the one most users want — it talks to your actual receipts, your actual Google Sheet, and your actual spending history. ChatGPT and Claude.ai users connect to the remote server at mcp.expensebot.ai, which exposes 38 tools (36 ExpenseBot + 2 ChatGPT Deep Research shims). Both delivery methods are listed on the official MCP Registry.

What can my AI assistant do?

Once connected, you can ask in plain English. Some example prompts that all work today:

• "Find every Uber receipt I expensed last quarter."
• "Submit this receipt to ExpenseBot." (drag a photo into ChatGPT or Claude)
• "Scan my Gmail for new receipts from the last 30 days."
• "What were my top 5 software vendors in Q1?"
• "Generate a Schedule C report for 2025 and share it with my accountant."
• "I paid $80 cash for parking at the airport — log it."
• "Show me my profit and loss for 2026 year to date."

Behind the scenes, those prompts route to one of 38 tools — 36 ExpenseBot tools plus 2 ChatGPT Deep Research shims (search + fetch). You don't call these by name; your AI assistant picks the right one based on what you ask.

You don't call these tools by name — Claude picks the right one based on what you ask. The list is here so you know what's possible. Power users link the MCP from their freelancer workflow so their AI assistant becomes the front-end for bookkeeping.

Connect ExpenseBot to ChatGPT

ChatGPT users connect via Developer Mode — OpenAI's standard pre-verified flow. The one-time informational warning disappears once the verified-connector approval lands (review currently pending).

1. Open ChatGPT → Settings → Apps & Connectors → Developer Mode
2. Click Add custom app and paste the server URL:

   https://mcp.expensebot.ai/mcp

3. Click Connect → Google login → Allow
4. 38 tools appear in your ChatGPT tool list. Try: "Find my receipt photos from last week and submit them to ExpenseBot"

Connect ExpenseBot to Claude.ai

Claude.ai Pro users connect via the Connectors panel — no config files, no npm. Dynamic Client Registration (DCR) means you leave the OAuth fields blank; Claude handles the handshake automatically.

1. Open Claude.ai → Settings → Connectors → Add custom connector
2. Server URL:

   https://mcp.expensebot.ai/mcp

   Leave all OAuth fields blank (DCR handles registration automatically).
3. Click Add → Connect → consent screen → Allow
4. Try it: attach a receipt photo and type "Submit this Uber receipt to ExpenseBot"

Already using a desktop AI client? Use the npm package below — same 35 tools plus CLI-friendly token rotation.

How to connect it (3 minutes) — Claude Desktop

1. Generate a token. In ExpenseBot, go to Settings → AI Assistant Tokens → Generate. Copy the token — it's shown once. Tokens are scoped to your account and can be revoked any time.
2. Add to Claude Desktop config. Open the config file in any text editor:
   • Mac: ~/Library/Application Support/Claude/claude_desktop_config.json
   • Windows: %APPDATA%\Claude\claude_desktop_config.json

   Paste this block (replace the token):

   {
     "mcpServers": {
       "expensebot": {
         "command": "npx",
         "args": ["-y", "@expensebot/mcp-server-auth"],
         "env": {
           "EXPENSEBOT_TOKEN": "your-personal-access-token-here"
         }
       }
     }
   }

   Just want the read-only docs server? No token needed:

   {
     "mcpServers": {
       "expensebot-docs": {
         "command": "npx",
         "args": ["-y", "@expensebot/mcp-server"]
       }
     }
   }

3. Restart Claude Desktop and test. Fully quit and reopen Claude. Try "show me my top 5 software vendors last quarter". Claude will ask permission to run the get_spending_summary tool — approve once, and the integration is live.

Other MCP clients

The same npm package works with every MCP-compatible client — only the config location changes.

Claude Code (CLI)

One command — no JSON editing:

claude mcp add expensebot -- npx -y @expensebot/mcp-server-auth --token=YOUR_TOKEN_HERE

Add --scope user to make it available across every Claude Code project.

Cursor

Edit ~/.cursor/mcp.json:

{
  "mcpServers": {
    "expensebot": {
      "command": "npx",
      "args": ["-y", "@expensebot/mcp-server-auth", "--token=YOUR_TOKEN_HERE"]
    }
  }
}

In Cursor: Settings → Tools & Integrations → look for the green dot next to expensebot. In chat, switch to Agent mode and prompt as usual.

Continue, Cline, Windsurf, custom MCP clients

Same npx -y @expensebot/mcp-server-auth --token=... command — drop it into your client's MCP config block.

ChatGPT and Claude.ai web use the remote server at mcp.expensebot.ai — see setup instructions above ( ChatGPT, Claude.ai ).

Why use AI Assistant access?

• Faster than the dashboard for ad-hoc questions. "What did I spend at Costco last month?" is one sentence in Claude versus four clicks in the UI.
• No app switching. If you live in Claude or Cursor for writing and code, your expense answers come to you.
• Conversational receipt submission. Drag a photo into chat, say what it was for, and it lands as a row in your sheet — categorized, with vendor and total extracted.
• Complex analytics in natural language. get_deep_analytics surfaces top vendors, anomalies, month-over-month deltas and unusual spend without you writing a single SQL query or pivot table.
• Cash entry by description. "$80 cash for airport parking yesterday" becomes a properly tagged cash expense with date, amount, category, and notes.

Security & token model

Personal access tokens (PATs) are the one credential the MCP server cares about. Each token is scoped to your account only — it can't read another user's data, even if leaked. Tokens have a configurable expiry you choose at issue time, and you can revoke any token from Settings → AI Assistant Tokens with immediate effect (the next request fails 401). Token issuance, use, and revocation are audit-logged in Firestore against your account so you can review what was used when. Tokens themselves are stored hashed (SHA-256) — even ExpenseBot staff can't read the original string out of the database.

Your underlying expense data never moves. Receipts stay in your Google Drive, your sheet stays in your Drive, and each MCP tool call reads or writes those files using your existing OAuth grants. The only data that flows through your AI provider is the specific tool result your assistant asked for — same trust model as any LLM tool-use call. If you're inviting an accountant later, the share happens through Google Drive sharing, not through the MCP token.

Compatible AI assistants

MCP support is growing quickly. Today the ExpenseBot MCP works out of the box with:

• ChatGPT — Plus/Pro plans. Connect via Developer Mode in Apps & Connectors. (setup ↑)
• Claude.ai — Pro plan. Connect via Settings → Connectors. (setup ↑)
• Claude Desktop — Mac, Windows, Linux. Native MCP support via npm package. (setup ↑)
• Cursor IDE — Mac, Windows, Linux. Add via Settings → Tools & Integrations. (setup ↑)
• Claude Code (CLI) — one command. (setup ↑)
• Continue, Cline, Windsurf — same npm package, different config path. (setup ↑)
• Custom agents built on the Anthropic, OpenAI, or open-source MCP SDKs.

If your client follows the MCP wire format (JSON-RPC over stdio), ExpenseBot works. The official and constantly-updated client list lives at modelcontextprotocol.io/clients.

Why this matters

Most expense apps see receipts. The ExpenseBot MCP lets your AI act on them — search, submit, scan Gmail, generate a tax report, share it with your accountant — using your actual Google Sheet as the source of truth. There's no separate mirror database to drift, no proprietary ledger to escape from, no "export to CSV" step. Your AI assistant reads the same sheet you read, writes the same rows you'd write, and triggers the same Gmail scanner you'd trigger from the dashboard.

That's the Gmail-Sheets-native angle. ExpenseBot was built from day one around the data your life already runs on — receipts in Gmail, ledger in Google Sheets, files in Drive — instead of on a closed app database. The MCP server inherits that. When you ask Claude to scan Gmail for new receipts, it triggers the same scanner that runs nightly. When it submits a receipt, the row lands in the sheet you already own. We're listed on the official MCP Registry at io.github.TotesMagotes/mcp-server-auth so registry-aware clients can find us automatically.

See pricing plans for full feature breakdown. The MCP servers are public on npm; account-scoped tools require an active ExpenseBot account.

New to the AI-first workflow? The tutorial walkthrough Manage Expenses from Inside ChatGPT or Claude shows real prompts, the 5-minute connector setup for ChatGPT and 3-minute setup for Claude.ai, the 5 workflows that replace separate apps, and the OAuth security model.

Frequently asked questions