What the Spendesk MCP cannot do — and what to do instead

Why this page exists

The Spendesk MCP has 62 tools: 36 that read and 26 that act. That covers a lot of finance work — spend analysis, invoice follow-up, month-end close, purchase orders, supplier maintenance. It does not cover everything, and the gaps are not always where you expect them.

An assistant that lacks a tool does not always say so clearly. It may answer a nearby question instead, or return an empty list that looks like "nothing to report". Knowing the limits upfront is the quickest way to trust the answers you do get.

The six things there is no tool for

Two of these catch people out more than the others.

"Mark as ready" is not an approval. The mark_payable_as_ready tool marks a payable as ready for accounting export. It is a bookkeeping step. It approves nothing and pays nothing.

Request tools cover card requests only. get_requests, get_request_by_id and spendesk_analyze_requests return single-purchase, subscription and multi-use card requests, and physical card loads. Travel budgets, purchase order requests, expense claims and mileage allowances are left out, with no marker and no error — so a "pending requests" total is a total of card requests.

What it looks like in practice

Ask for something there is no tool for, and a well-instructed assistant says so and offers the nearest useful thing.

The same goes for budgets. The assistant analyses spend; it cannot compare it with a budget it has never seen.

Limits that are about access, not tools

Some things fail for a reason that has nothing to do with the tool list.

• Role. Only Controllers and Account Owners can use the assistant, for reading as for acting. Admins can set up the connection but cannot use it; Requesters cannot connect.
• Write permissions are off by default. Write actions are in Beta. An Admin or Account Owner enables each write permission on the connection, then each user disconnects, reconnects and ticks it. A tool whose permission is not ticked looks exactly like a tool that does not exist.
• One company at a time. On an organisation-level connection the assistant works through companies one by one. Tell it which entity you mean.
• Different currencies are not mixed blindly. Spend analyses give totals in the company's functional currency. Lists of invoices keep each invoice in its own currency, and those amounts are not added together. Across entities in different currencies there is no group total: the MCP supplies no exchange rates.

Limits inside the tools that do exist

A tool can exist and still not answer the question you had in mind.

• A past reference date is not a historical aging. Asking for overdue invoices "as of 30 June" re-ages the invoices still unpaid today against that date. It does not rebuild the June month-end position.
• For a payable, a failed export and a pending one look the same. The assistant can tell that a payable is not exported yet; it cannot tell whether an export was attempted and failed.
• Wallet balance is recent, not real-time. The balance comes from a short cache, and wallet load history can lag by up to a day for an identical query.
• Some actions cannot be undone. Cancelling or closing a purchase order, committing an accounting export, and marking a payable or a settlement as exported have no reverse tool. Keep a human confirmation on each one.

Key takeaways

• There is no tool to approve, pay, read budgets, send alerts, or create or change cards, users and cost centers — use Spendesk itself for these
• For budget vs. actual, supply the budget; the assistant supplies the actuals
• Request tools cover card requests only: travel budgets, expense claims and mileage are not in them
• A missing tool is often a missing permission: check the role and the ticked write permissions before concluding it does not exist
• The authoritative list is the MCP tool reference, and the limits are listed in Connect an AI assistant