The Model Context Protocol (MCP) is related to generative AI (GenAI), and it is something those in accounting should become familiar with. MCP makes it easier for GenAI solutions to work with your applications, online services, and computer.
Imagine asking your computer to pull up next quarter’s cash flow forecast, to check it against the latest supplier invoices, and to prepare a presentation, and it just does it. No tab switching. No data exports. No reformatting. Just results. This scenario is becoming increasingly plausible because of a new technical standard called the Model Context Protocol (MCP). While MCP might sound deeply technical, its implications for accountants are anything but abstract.
MCP is a relatively new technical term related to generative AI (GenAI), but it is one that those in accounting should become familiar with. Sometimes it is likened to the universal serial bus (USB): just as USBs made it easier to plug different devices into your computer or to charge things, MCP makes it easier for today’s GenAI solutions to work with your applications, online services, and your own computer. Since MCP’s introduction in November 2024, it has received enormous attention.
As AI – especially AI agents – become commonplace and are integrated into our daily activities, secure and standardized ways to provide interoperability between tools and agents has been on the rise. This blog introduces MCP, how to use it, and how it will impact accountants and auditors.
The Rise of Agentic AI and the Role of MCP
The accounting profession is entering a new era where GenAI doesn’t just assist, it acts. ChatGPT, Claude, Gemini, and others are rapidly evolving from passive assistants to proactive agents that interact with business tools to complete complex workflows autonomously.
While AI agents have shown some ability to interact with applications on their own, Anthropic – the company best known for the Claude GenAI solution – determined that standardizing the instructions for integrating large language models (LLMs) with external data sources and tools would provide a more efficient and scalable way to help any AI understand the functions and capabilities of resources to more efficiently, safely, and effectively create an interoperable agentic AI ecosystem. Thus, MCP was born.
With MCP, agents don’t need special plugins or hardcoded integrations. They can simply query what functions are available and interact with it, similar to how humans open a file or run an application.
Anthropic introduced MCP in November 2024, and vendors, solution providers, and developers began to publish MCP solutions. But interest in MCP really took off in May and June 2025 when Google, OpenAI, and Microsoft announced that native MCP support would become standardized, which was followed by a wave of integrations from cloud platforms, enterprise tools, and AI frameworks.
For CPAs, MCP is quickly becoming something you cannot ignore. It will impact how you manage your firm, advise your clients, deliver assurance, and protect data.
A Nontechnical MCP Primer
MCP is an open protocol that lets AI agents interact with computer resources (files, databases, apps, even external tools) in a structured, secure, and consistent way. Think of it as a universal translator between AI’s natural language understanding and the functional reality of your business systems.
It works using a client-server model. It’s not quite like the old days of networking, where there was a file server and your computer was the client. Instead, MCP is more like this:
- The MCP server advertises what actions are available (e.g., “generate trial balance,” “submit 941 form”).
- The MCP client discovers those capabilities and uses them appropriately. The client interprets the requests from the LLM, translates them into the messages the MCP server expects, and then translates the responses back to the LLM.
What makes MCP different is that these capabilities are described in human-readable, AI-usable metadata, meaning the agent doesn’t need hardcoded integration. It can discover what your applications (like your accounting software) can do and then starts using them. For example, with Xero accounting software, the MCP exposes more than 40 functions from list-accounts (retrieve a list of accounts) to update-payroll-timesheet-line (to update an existing line on a payroll time sheet).
For accountants, this means reducing the friction of integrating legacy systems, cloud services, and third-party tools into their workflows.
Why May/June 2025 Was a Turning Point
When Microsoft confirmed in May 2025 that Windows 11 would natively support MCP, allowing agents to discover and interact with local apps, services, and files, within just a few weeks the following occurred:
- Anthropic launched an update to Claude for Desktop, allowing Claude AI to act on files and apps using MCP.
- MCP Superassistant was released, adding “built-in” MCP discovery to GenAI tools accessible in Chrome.
- Enterprise tools emerged or expanded with support for SAP and QuickBooks, and other key players began publishing their own MCP interfaces.
Why CPAs Should Care
Some might say it is still early. Using the MCP support in Claude Desktop or adding it to ChatGPT with Superassistant requires some technical skills, such as installing and loading Python programs and working with files from repositories like Github. Plus, the Windows 11 launch is still only an “early preview of the MCP platform capabilities to developers in the coming months for the purposes of feedback.”
However, it is the potential benefits of safely streamlining AI access to the tools your practice or business already uses that make MCP so exciting:
- Efficiency – AI agents from different solutions can automate reconciliations, generate reports, prepare filings, and cross-check data across tools using MCP, without custom scripts. Imagine telling your AI assistant to update client records, run comparative financials, and summarize key variances, even across different apps.
- Client Service Transformation – With tools like Claude for Desktop, you can empower clients to self-serve audit prep, tax documentation, or financial forecasting.
- New Assurance Frontiers – As MCP-enabled agents begin to make decisions and take actions, CPAs will be needed to verify integrity, trace interactions, and offer assurance on AI-enabled workflows.
- Risk Management – Improperly configured MCP servers or untrusted agents pose real risks, including data leakage, process errors, and noncompliance. Vetting and certifying MCP servers is a long way off. Repositories like MCP.so may currently list 15,000-plus MCP servers, but there’s been no vetting that they will work as promised.
Risk and Control Issues
With rapid advancement and adoption comes risks but also opportunities to advise clients. Some areas CPAs must keep an eye on include the following:
- Over-Permissioned Tokens – Some early MCP implementations issue credentials that give agents too much power, allowing them to modify or access unrelated systems.
- Missing Audit Trails – Without enforced logging, it is unclear what agent took what action, when, and why.
- Ambiguous Source Trust – Not all MCP servers are equal. A rogue server could mislead the agent.
- Approval Workflow Gaps – MCP does not enforce “human-in-the-loop” for high-risk actions unless specifically configured.
- Agent-Generated MCPs – Advanced agents may create their own micro-interfaces for efficiency, raising the question of who approves those workflows?
The Bigger Picture
MCP is just the plumbing. The broader vision is a world of interactive agent ecosystems. MCP is an important player in the new AI agentic “stack.” With the simplification and scalability MCP can bring, we will soon see:
- Multiple agents collaborate, delegate, and negotiate tasks. (Another key part of the stack is Google’s Agent-to-Agent (A2A), which can facilitate this step.)
- Tools register themselves with MCP servers.
- Agents synthesize new capabilities by composing lower-level MCP actions.
- Registries emerge that catalog public and private MCP services.
In this world, trust boundaries blur and controls become code. CPAs who specialize in this area can help set boundaries, define acceptable agent behavior, monitor activity, and provide assurance on outcomes.
Practical Steps for CPAs
So, what are your next steps? First, you will want to track the tools your company or practice use or support to monitor whether MCP servers have been published. As they increase, you’ll want to make sure you and your clients have governance policies in place regarding publishing and using MCP, and under which conditions. That will include logging, evaluating, and tracking MCP servers in use. You may want to consider building SOC-style services or MCP audit frameworks.
From Edge Case to Core Competency
MCP has moved from obscure innovation to foundational infrastructure in less than six months. By providing a unified interface for AI agents to interact with computer resources, MCP empowers professionals to automate tasks, integrate systems, and work smarter. It will define how AI agents interact with business systems, data, and workflows going forward. For CPAs, that means new efficiencies, new services, and new responsibilities with fewer silos, leading to opportunities to provide trust around autonomous systems.
Whether you create MCP servers, use them, or audit them, you are now part of this ecosystem. CPAs don’t need to write the code, but they do need to understand the impact. MCP is coming to a desktop near you. It’s time to pay attention.
Eric E. Cohen, CPA, is owner and chief AI officer of Cohen Computer Consulting in Mechanicsburg, Pa. He can be reached at eric.e.cohen@computercpa.com.
Sign up for PICPA's weekly professional and technical updates by completing this form.
Statements of fact and opinion are the authors’ responsibility alone and do not imply an opinion on the part of the PICPA's officers or members. The information contained herein does not constitute accounting, legal, or professional advice. For actionable advice, you must engage or consult with a qualified professional.
