Drive your Spendesk finance ops with Claude, ChatGPT, and Dust
A library of skills and agents built for finance teams — CFO, Controller, AP Manager, FP&A Analyst, Procurement. Works with Claude Code, ChatGPT, and Dust.
New to the Spendesk MCP? Take the guided tour
A skill that calls real tools on your own data, in under two minutes, then recommends what to do next.
Write tools are available in Beta
The MCP is moving from analysis to action: 26 write tools are available in Beta (suppliers, payables, purchase orders, chart of accounts, accounting exports, analytical fields). Their permissions are off by default — here are four examples.
archive_supplierArchives or reactivates a supplier.
create_suppliersCreates a new supplier.
mark_payable_as_readyValidates a payable's accounting information and marks it ready for export.
update_payableUpdates a payable's accounting fields (expense/tax account, accounting date, cost center, analytical fields, amortisation).
In Beta: an Admin or Account Owner enables each write permission on the connection, then each user reconnects and ticks it. Some actions cannot be undone.
What protects you
The assistant acts on your behalf, within Spendesk's controls. Here is what Spendesk enforces, and what is an instruction given to the assistant.
Your permissions
Enforced by SpendeskThe assistant only sees the tools you ticked when connecting. Write permissions are unticked by default.
Your role
Enforced by SpendeskEvery call checks your role in the company concerned: only Controllers and Account Owners can use it.
No payment, no approval
Enforced by SpendeskThere is no tool to pay, schedule a payment, approve or reject. Those actions stay in Spendesk.
An action log
Enforced by SpendeskEvery call the assistant makes is recorded in the connection's action log, which Admins and Account Owners can open.
A confirmation before acting
Instruction to the assistantThe assistant is instructed to show you each change and wait for your go-ahead. It is an instruction, not a lock: keep your assistant's own approval on for write actions.
One concrete example per role
The same need, asked to Claude or to Dust — both platforms call the same Spendesk MCP tools.
Multi-entity cash flow, supplier concentration risk, cash flow forecasting
Here's your consolidated treasury position across 3 entities:
Total wallet balance: €842,300
Spend over the last 30 days: €284,600 (-6% vs previous month)
Want the per-entity breakdown, or a 4-week projection?
Month-end close, GL mapping, VAT audit, duplicate invoice detection
Check for duplicate invoices this month
I analyzed 218 invoices for July 2026. I found 1 potential duplicate:
Duplicate — Confidence: 91%
Want me to open both invoices to compare attachments?
Invoice aging, payment priority, suppliers without IBAN
3 invoices are overdue, totaling €12,400:
🔴 Sopra Steria — €6,200 (12 days)
🔴 Capgemini — €4,100 (8 days)
🟡 Docusign — €2,100 (3 days)
Want me to prioritize this week's payment batch?
Spend trends, SaaS growth tracking, cost center analysis
Compare spend by cost center over the last 3 months
Cost center spend comparison (May → July 2026):
| Cost center | May | June | July | Trend |
|---|---|---|---|---|
| Engineering | €78,200 | €84,100 | €91,200 | ↑ +16% |
| Marketing | €42,000 | €38,600 | €45,300 | ↑ +8% |
| Sales | €29,400 | €31,000 | €27,800 | ↓ -5% |
Engineering is the only center trending up for 3 consecutive months.
Invoice-PO matching, tail spend, purchase request aging
5 purchase orders are open with no matching invoice, totaling €34,200:
Want me to follow up with suppliers on the 2 oldest ones?
Install a skill in 30 seconds
Download the .md file
Place in .claude/ of your project
Use in natural language in Claude Code
Frequently asked questions
What is a finance ops MCP?
A finance ops MCP (Model Context Protocol) is a server that connects an AI assistant like Claude, ChatGPT, or Dust directly to a company's finance data — invoices, suppliers, purchase orders, treasury, cards — so it can answer questions or automate tasks without manually exporting files.
Does Spendesk have an MCP server?
Yes. The Spendesk MCP exposes 62 tools today: 36 read tools (invoices, suppliers, purchase orders, cards, treasury, cost centers, transactions) and 26 write tools in Beta (suppliers, payables, purchase orders, chart of accounts, accounting exports, analytical fields), whose permissions are off by default — usable from Claude, ChatGPT, or Dust.
What's the difference between the Spendesk MCP and this skills library?
The Spendesk MCP provides the raw tools (API calls). This library provides 75+ ready-to-use skills — tested instructions that combine those tools to answer a specific business need (PO accrual, supplier review, monthly close...).
Does the Spendesk MCP work with Claude, ChatGPT, and Dust?
Yes, all three platforms call the same Spendesk MCP tools. Every skill in the library specifies how to install it on Claude Code or on Dust; ChatGPT connects via a client ID/secret issued in the Spendesk app.
Do I need IT to use it?
No development is needed. An Admin or Account Owner creates the connection once in Spendesk (Settings → Integrations → MCP) and passes its details to the administrator of your Claude, ChatGPT or Dust workspace. After that, each user signs in with their own Spendesk account and approves their own access.
Can the assistant change or pay something without me?
It cannot pay or approve anything: no tool does that. Write actions (creating a purchase order, updating a supplier, generating an accounting export) are in Beta and off by default; once enabled, the assistant is instructed to show you each change and wait for your go-ahead.
Who can use the Spendesk MCP?
Controllers and Account Owners, for reading as for acting. Admins can set up the connection but cannot use it, and Requesters cannot connect.
Can spend be compared with a budget?
Not directly: the Spendesk MCP does not read budgets. You supply your budget figures, typed in the request or attached as a file, and the assistant compares them with actual spend from Spendesk.
What is the difference between a skill and simply asking the assistant?
A question returns data. A skill carries the procedure: which checks to run, in what order, and the known traps of each tool — for example never adding different currencies together, or knowing that an expense not yet prepared is left out of an export without warning.