Get FP&A best practices, research reports, and more delivered to your inbox.
Bottom line: an MCP server is good when it can do anything a human can do in that software — not just read a few summary endpoints. That single test, put forward by Aleph co-founder and CEO Albert Gozzi during our Claude Skills for Finance webinar, separates the platforms an agent can genuinely work in from the ones that merely list MCP on a feature page.
MCP has gone from an unfamiliar acronym to a buying criterion in about a year. The honest version of the buying advice is stronger than most vendors will put in writing: it is hard to justify purchasing a finance system today that has no MCP at all, because it caps how much of your work an agent will ever be able to touch. But presence is not quality, and the gap between the two is where teams get disappointed.
What is an MCP, in finance terms?
MCP — the Model Context Protocol — is what lets an application talk to Claude or ChatGPT. That is genuinely all it is. If you would normally open your ERP, navigate to the reports section, pick an income statement and download it, an MCP is the same journey described in a way an agent can follow.
What makes it matter is that it is a common language rather than a bespoke pipe. Wiring two systems together through their APIs has historically meant paying integration consultants somewhere in the tens of thousands. MCP replaces that per-pair custom work with one protocol every tool can implement, which is why the ecosystem expanded so quickly once the major assistants supported it. The protocol itself is public — the Model Context Protocol specification is the canonical reference if you want the technical detail.
For the wider picture of how this fits a finance stack, see our MCP guide for finance teams.
What makes an MCP good rather than just present?
Use the human-parity test: ask what a person can do in the software, then ask how much of that the MCP exposes. Six checks separate a real implementation from a checkbox.
The permissions row is the one finance teams underestimate. If an MCP runs on a single shared service account, then anyone who can reach the agent can reach everything the account can reach — which is the opposite of what you want when a department head asks about their own spend line. Per-user permission inheritance is what makes agent access safe to roll out beyond finance.
What MCPs still cannot do
Two limits are worth knowing before you design a workflow around one.
First, the protocol does not carry files the way people assume. Passing a PDF or an image through MCP is not currently supported, so a workflow that depends on handing a document from one system to another through an agent will not work the way you sketched it on a whiteboard.
Second, coverage is uneven. Where a system has no MCP at all, the fallback is to have the agent drive a browser and download a CSV. That works and it is a legitimate stopgap, but it is brittle: it breaks when the page layout changes, and it gives you no traceability. Treat it as a bridge, not a design.
MCP, API and integration are not the same thing
These get used interchangeably and it causes real confusion in vendor calls, so it is worth being precise. An API is a system's programmatic interface. A custom integration is code somebody wrote to join two specific systems together. MCP is a shared protocol that lets any compliant agent talk to any compliant system.
The economic difference is the reason MCP spread so fast. Custom integration work scales with the number of pairs: connecting five systems to each other is ten separate projects. A protocol collapses that to one implementation per system. If a vendor tells you they will build you a custom connector instead, you are being offered the expensive version of a solved problem.
What a finance-grade MCP needs on top of the basics
The generic checklist above applies to any software. Finance adds three requirements that general-purpose guidance misses, and they are the ones that decide whether an agent's output survives contact with an auditor.
Period awareness. The MCP should expose whether a period is open, closed or adjusted, and an agent should be able to tell the difference. An agent that reads an open period and reports it as final will produce a confidently wrong number every month, and nothing about the output will look wrong.
Dimension fidelity. Your entity, department, cost-centre and class structures need to come through named as you name them. If the MCP flattens them, every agent query has to re-derive the mapping, which is both slow and a place for silent error.
An audit path. Any figure an agent reports should be traceable to the records behind it. This is the difference between an answer you can take into a board meeting and one you have to go re-derive by hand before you dare use it — the distinction we draw out in explainable AI for FP&A.
Which finance platforms have an MCP?
Coverage now spans FP&A, ERP and the warehouse layer, though quality varies more than availability does.
Availability reflects public vendor documentation as of August 2026; confirm current state and the depth of each implementation directly with any vendor. For the FP&A layer specifically, we keep a fuller list in MCP-compatible FP&A platforms and FP&A platforms that work with Claude.
How to test an MCP before you buy
Do not accept a demo on the vendor's dataset. Ask for four things on yours:
- A write, not just a read. Have the agent change something and show you the record it changed.
- A line-item drill. Ask for a number, then ask which transactions make it up. If it cannot get below a summary, your commentary will never be defensible.
- Your own dimensions. Ask for a report cut by the dimensions you actually report on, named the way you name them.
- Two different users. Run the same question as a finance user and a department head, and confirm the answers differ appropriately.
That last test is the one that reveals whether permissions are real. It also mirrors how we think about agent access generally, covered in AI accuracy and auditability in FP&A.
Where Aleph fits
Aleph's MCP exposes the reporting and planning layer — actuals, budgets, forecasts and the dimensions you report on — to Claude and other agents, with each user's own permissions and a drill path back to source. The point is not that the agent is smarter; it is that it is working from your chart of accounts, your metric definitions and your plan rather than guessing at them.
See it in the Aleph Agent, or watch Albert build a skill and connect it to the Aleph MCP live in the Claude Skills for Finance recording. The Claude skills for finance guide includes six ready-made .skill files to start from.
Get FP&A best practices, research reports, and more delivered to your inbox.
