Practice Management MCP: Vendor vs Custom vs Neither

By Jude Lee · · News

Two accounting firm staff reviewing job status data on a monitor while discussing workflow

The news: practice management data is showing up inside AI assistants

Practice management vendor Financial Cents has announced an MCP server that lets firm data be queried from everyday AI assistants. Confirm the announcement and its current scope in the vendor’s own newsroom and product documentation before you rely on it — vendor press releases get revised, and feature availability often differs by plan.

Expect more of this generally. MCP — the Model Context Protocol, an open standard introduced by Anthropic in late 2024 for giving AI assistants governed access to tools and data — has moved from a developer curiosity to something practice management, ledger, and document vendors ship as a feature.

As of 2026 the landscape is moving monthly, so verify current availability in each vendor’s official documentation rather than trusting a blog post (including this one).

What an MCP server actually is, in firm terms

Strip the acronym away and an MCP server is a controlled doorway. It publishes a list of things an AI assistant is allowed to do — “list open jobs,” “get client details,” “show time entries for a period” — and the assistant can call those, with your credentials and your permissions, during a conversation.

Two honest clarifications. First, it is not a data dump: the assistant sees what the server exposes, nothing more. Second, it is not automation by itself. An MCP server gives an AI access; whether anything useful happens depends on whether you’ve defined the workflow, the review gate, and the output format. We’ve covered the plumbing distinction between a raw API and an MCP layer in practice management APIs vs MCP servers.

Where it fits a real firm workflow

A concrete tax-season example. It’s March 20. The manager wants to know: which 1040 jobs are sitting in “waiting on client” with no contact logged in the last seven days, and which of those have a signed engagement letter but no retainer?

Without MCP, that’s three exports and a VLOOKUP, or a saved report that answers 70% of it. With a practice management MCP server connected to an assistant, it’s a question — and the follow-up questions are free: “group those by preparer,” “draft a chase email for each, in our tone, for my review.”

Now the failure mode, because it is specific and it will happen to you. A partner asks for every return due in the next ten days. The exposed tool filters on a due-date field, and eleven jobs in the firm have a null due date because they were rolled forward before the template was fixed. Those eleven are silently dropped from the result set, and the assistant answers with complete confidence — no warning, no footnote. A second version of the same problem: a job whose status field was never moved off “in review” after delivery gets reported as outstanding, and someone chases a client who already paid. Neither is an AI malfunction. Both are your data hygiene, surfaced faster and with more authority than a spreadsheet would have claimed.

That is also why the drafted action matters more than the read. Reading data is mildly useful; turning the read into a drafted action that a human approves is where the hours actually move — see our take on building AI review gates rather than autopilot.

Access without a review gate isn’t automation. It’s a faster way to be confidently wrong in front of a client.

Vendor MCP server vs building your own

Vendor-provided MCP server

Good when: the questions you care about live in one system; you want value this quarter; you have no internal dev capacity. Setup is usually an auth flow and a config file. The vendor maintains it as their schema changes.

Cost shape: a per-seat or per-firm subscription line you don’t control — the vendor sets the price, the plan tier it sits in, and when it changes.

Limits: you get the tools they chose to expose. Scope granularity, audit logging depth, and multi-client isolation are whatever the vendor built. Cross-system questions (“jobs in PM vs. unbilled WIP in billing vs. open queries in the ledger”) stay unanswered.

Custom MCP server

Good when: the answer requires joining two or three systems; you need field-level least privilege; you want your own audit trail of every call an assistant made; or you’re packaging a repeatable job as a skill the whole team runs identically.

Cost shape: build hours up front, then ongoing maintenance — someone has to fix it when an upstream API version changes, which is not a one-time cost.

Limits: it is software. Someone owns it, versions it, and patches it. For a small firm with one system of record, that’s overhead you may not need.

And a third option that gets skipped: neither. If the question is answered by a saved report emailed every Monday, build the saved report. Rules-based automation and plain reporting are frequently the cheaper, more reliable answer — we laid out that triage in rules, AI agents, or neither.

Ten minutes of due diligence before you connect client data

  1. Confirm read-only first

    Start with a server (or scope) that cannot create, edit, or delete. Write access is a separate decision with separate controls.

  2. Read the scope list literally

    Ask the vendor exactly which objects and fields are exposed. “Client data” is not an answer; “client name, entity type, job status, due date” is.

  3. Check whose permissions apply

    Does the connection inherit the signed-in user’s role, or does it run as an admin service account? The second is a bigger blast radius.

  4. Ask where the text goes

    Which model provider processes the content, under what retention and training terms? Your answer shapes what you’re permitted to send.

  5. Demand an audit log

    You want a record of who asked what, which tools were called, and when — for quality management and for incident response.

  6. Test with a dirty dataset

    Ask questions where you already know the answer, including edge cases: inactive clients, jobs with null due dates, duplicates.

  7. Define the review gate

    Nothing goes to a client, a taxing authority, or the ledger without a named human approving it.

  8. Write it into your security plan

    The IRS’s Publication 4557, Safeguarding Taxpayer Data, directs tax professionals to maintain a written data security plan; a new AI connection to client records belongs in it.

  9. Check confidentiality obligations

    The AICPA Code of Professional Conduct includes rules on confidential client information; confirm your state board’s requirements before routing client data to a new processor.

  10. Pilot one workflow, then expand

    Resist the urge to connect everything at once.

What this means for accounting jobs

The anxious framing of this question — whether AI is coming for CPA jobs — misses the point. What MCP access removes is the retrieval layer: exporting, filtering, reconciling two lists, re-keying. What it does not remove is judgment, client relationships, professional skepticism, or responsibility for the file. A model that confidently tells you a job is complete because a status field was never updated has not replaced anyone; it has just surfaced your data hygiene problem faster.

The practical shift is in who does what. Firms adding “accounting automation specialist”-type responsibilities are usually reassigning an existing senior who knows the workflows, not hiring a data scientist. The scarce skill is specifying a job precisely enough that an agent can repeat it — closer to writing a very good SOP than to coding.

How to rank tools for your own firm

There is no ranking that survives contact with your firm, but there are honest criteria: does it connect to your system of record, can you scope access narrowly, does it log what happened, and can your team actually operate it on a Tuesday in March. A vendor MCP server from the practice management tool you already pay for will beat a more capable custom build you never finish. Equally, no MCP server fixes a chart of accounts nobody maintains or a job board half the team ignores.

If you’re deciding how client data may be handled at all before you compare tools, start with the governance question — our breakdown of four setups for using AI with client data is the better first read.

Modeling the payoff without making numbers up

Don’t accept a vendor’s hour-savings headline, and don’t accept ours — we don’t have one. Build the estimate yourself:

If the arithmetic only works when you assume the agent runs unreviewed, the business case isn’t there yet.

Not sure where to start?

Get a free automation audit: we map your bookkeeping, month-end close, client onboarding, document collection, and AP/AR — and show you what's worth automating before you spend a dollar.

Get a free automation audit

Related guides