Our Own Books, Run Through Claude - A Custom Xero Connector

In August 2026 we moved our books into Xero from another accounting platform. That meant the full history: customers, products, repeating invoices, every past invoice with its original number, the payments against them and months of coded bank transactions. On top of that, three Stripe accounts bill into the same ledger, and we look after a second Xero organisation as well.

We wanted to tell Claude what needed doing in plain English, and have it done properly, in the right organisation, every time.

Xero now offers a free read-only connector inside Claude and ChatGPT, and an official MCP server that can create and update records. Both are worth using. But the connector can't change anything, and the MCP server only exposes the parts of Xero's API its makers chose to wrap. There are no tools to void or clean up records, it works with one organisation per connection, and it has no way to load hundreds of records and pick up where it left off.

For a migration and day-to-day bookkeeping across two organisations, that wasn't enough. One wrong setting, and real entries would land in the wrong set of books.

A Xero toolkit that Claude runs on our behalf, with the whole Xero Accounting API open to it rather than a short list of tools. Every job is written to be safe to repeat, can be previewed before it posts, and refuses to run unless it is pointed at the organisation it was built for. Claude reads the request, picks or writes the job, shows the dry run, and runs it once it has been checked.

During the migration, a stray setting on one machine quietly pointed requests at the wrong organisation. Nothing like that can happen now. Every job checks three things before it sends a single request: the organisation is one it knows, the login is authorised for that exact organisation, and the organisation's live name matches. Most jobs are also pinned to the one ledger they were written for and refuse to run anywhere else.

• 186 historic invoices, with their original numbers and dates, totals matching the old system to the cent.
• 147 customer payments, allocated so debtor balances tie out.
• 45 contacts, 19 repeating invoice templates and the full chart of accounts with tracking.
• 528 bank transactions, loaded across two days because Xero caps each organisation at 5,000 calls a day. The job simply skipped what had already landed and finished the rest.

• Xero API: any Accounting API endpoint plus fixed assets and PDFs, through one custom connection per organisation, with read-only scopes for reporting.
• Claude: plans the job, writes or picks the script, shows the dry run and runs it once approved.
• AWS: the daily Stripe sync runs as a scheduled serverless job; everything else runs on demand.
• Rate limits: automatic back-off when Xero says slow down, and resumable jobs for the daily cap.

Our books are kept by asking. Raising a one-off invoice, fixing a contact, stopping a retainer or checking what hasn't been billed is a sentence to Claude, checked before it posts, in the right organisation every time.