Skip to Content
WonderMoney Docs
Settings, Security & Privacy

Connect an AI client

Let Claude Desktop, Cursor or another MCP client ask about your own finances with a read-only token.

WonderMoney runs a read-only Model Context Protocol (MCP) server. An MCP client such as Claude Desktop, Cursor or VS Code can ask it about your accounts, spending, budgets, goals, net worth and portfolio, and answer in the client's own chat.

The server is at POST /v1/mcp on your API. It is not a separate account and it has no write tools: it can read and, if you allow it, propose a change for you to confirm. It never changes anything itself.

What the client can see

The client sees exactly what you see:

  • The household's shared accounts, transactions, budgets, categories, goals and portfolios.
  • Your own private accounts and the private accounts shared with you, and not a household member's private accounts that are hidden from you.
  • Never another household's data.

Every tool takes your identity from the token, so the same privacy rules that apply in the app apply here.

What the client can do

The client can call these read-only tools:

  • get_account_balance_summary, list_accounts, get_net_worth_summary
  • get_cashflow_summary, get_spending_by_category, get_spending_trend, get_large_transactions_summary
  • search_transactions (date range, account, category, amount range, up to 100 rows)
  • get_budget_status, compare_budget_periods, list_budgets, list_categories
  • get_goal_progress_summary, get_recurring_payments_summary
  • get_portfolio_summary, get_portfolio_income_summary
  • get_insights

It cannot create, change or delete anything, and it cannot reach the household, bank connections, exports or token management.

Proposals

With the proposals scope (mcp:propose), the client can also call propose_categorize_transactions, propose_mark_transactions_reviewed and propose_update_transaction_note. Each one checks that every transaction is one you can see and drafts a proposal; it does not change the transaction. The proposal waits in Settings > Assistant > Pending changes (and More > Pending changes on mobile) until you apply or dismiss it, and expires after seven days. The client cannot apply it, and no API token can. See Assistant settings.

Create a token

  1. Open Settings → API tokens.
  2. Under Connect an AI client, select Create an AI client token.
  3. Copy the token immediately. It is shown once.

The shortcut creates a token named "AI client" with the AI client (MCP) scope and a 90-day expiry. If you use two-factor authentication, enter your authenticator code first. Tick Let the client propose changes for me to confirm first to add the proposals scope. You can also create the token by hand in the form above, choosing the same scopes.

The mcp:read and mcp:propose scopes reach only the MCP endpoint: mcp:read exposes the read-only tools and mcp:propose the proposal tools. It does not grant the ordinary data routes, and a read token does not reach MCP: the two are separate credentials on purpose.

Client configuration

Claude Desktop bridges a remote MCP server with mcp-remote. Add this to its claude_desktop_config.json:

{
  "mcpServers": {
    "wondermoney": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://api.example.com/v1/mcp",
        "--header",
        "Authorization: Bearer <your-token>"
      ]
    }
  }
}

Cursor and other clients that speak Streamable HTTP directly use the URL and header form:

{
  "mcpServers": {
    "wondermoney": {
      "url": "https://api.example.com/v1/mcp",
      "headers": {
        "Authorization": "Bearer <your-token>"
      }
    }
  }
}

Replace the API address with your own and <your-token> with the value you copied. The endpoint is stateless: the client sends the token on every request, and there is no session to keep.

Audit log

Every tool call is written to the assistant audit log with the tool name, the token and the time. Amounts and descriptions are never written there. Settings → API tokens → Connect an AI client lists the most recent calls.

Revoke access

Select Revoke next to the token in Settings → API tokens. Revocation is immediate: the next request from the client gets 401 Unauthorized. Delete the server entry from the client's configuration as well, so it stops retrying.

Rate limits

MCP traffic uses the token's own request budget, and each token may make a limited number of tool calls per minute (60 by default). A client over either limit gets an error and can retry shortly. The budget is separate from your browser session.

Self-hosting

The server is on by default. Set MCP_ENABLED=false on the API to turn it off: POST /v1/mcp then answers 404, and GET /v1/instance reports features.mcp: false, so the settings panel disappears. See Configure.

On this page