「会話もいらずにマクロが直る」エクセルの神様の答えにたどりついた話A developer shares how they built an MCP-based workflow that automatically…

A developer writing on Qiita has described building a workflow that uses the Model Context Protocol (MCP) to automatically diagnose and repair Excel macros, reportedly without the usual back-and-forth prompting that debugging with a chatbot normally requires. The account matters because spreadsheet automation written in VBA remains deeply embedded in business operations, yet maintaining and fixing that code is often a manual, error-prone chore handled by a small number of specialists. If an AI-driven pipeline can shorten that loop, it points to a broader shift in how legacy office automation might be maintained.

At the center of the approach is MCP, an open standard introduced by Anthropic in late 2024 to give large language models a consistent way to connect with external tools, files, and data sources. Rather than pasting code into a chat window and copying results back manually, MCP lets an application expose capabilities—reading a workbook, running a macro, capturing an error message—through a server that the model can call directly. In this case, the developer appears to have wired an MCP server to the Excel environment so that the language model can inspect the macro, observe what fails at runtime, and apply a corrected version, all as part of a single automated exchange.

The phrase in the original title, roughly "macros get fixed without even a conversation," captures the intended benefit. Traditional LLM-assisted debugging is interactive: the user reports a symptom, the model proposes a change, the user tests it, reports the new error, and the cycle repeats. By contrast, the described setup seems to close that feedback loop programmatically. The model receives the failure context directly from the MCP server, reasons about the cause, and pushes a fix back into the workbook, reducing the need for a human to mediate each step. The author frames this as arriving at the answer they had been searching for, suggesting earlier attempts relied on more manual trial and error.

Technically, VBA is a fitting target for this kind of automation. It is a mature, well-documented language, so language models tend to have substantial training exposure to its syntax and common failure patterns, such as type mismatches, out-of-range references, and broken object handling. Runtime errors in VBA also surface with specific numbers and descriptions, which gives an automated agent concrete signals to act on rather than vague guesses. That combination—predictable error reporting plus a language the model already understands well—likely makes Excel macros an easier proving ground than more idiosyncratic or sparsely documented codebases.

It is worth placing this within the wider MCP ecosystem, which has expanded quickly since the standard was published. A growing number of MCP servers now exist for filesystems, databases, version control systems like Git, and productivity tools, and major vendors including OpenAI and Google have signaled support for the protocol or compatible approaches. Microsoft has separately pursued AI integration inside Office through Copilot, which brings natural-language assistance into Excel, Word, and other applications. The Qiita author's project differs in spirit from that commercial route: it is a self-assembled, developer-controlled pipeline rather than a packaged feature, which appeals to users who want to customize behavior or keep processing within their own environment.

Some caveats apply. Automatically rewriting macros without human review carries risk, particularly when spreadsheets drive financial calculations or feed downstream systems, so validation and version control remain important safeguards. The reliability of such a workflow also depends heavily on the quality of the error context the MCP server exposes and on the specific model behind it; results are likely to vary with more complex or poorly structured macros. As a single practitioner's write-up rather than a benchmarked study, the account should be read as a demonstration of what is now feasible rather than a guarantee of consistent outcomes.

Still, the example illustrates a practical direction for MCP beyond chat interfaces: using the protocol to let models operate directly on real artifacts and act on live feedback. For teams that depend on aging VBA code but lack dedicated maintainers, an approach that automates the diagnose-and-fix cycle could meaningfully reduce friction, provided it is deployed with appropriate testing and oversight.