「会話するだけでマクロが直る」は「会話するだけですごいマクロが作れる」だった話By connecting Excel with Claude Code via MCP, the author discovered that…

An Excel user has documented how pairing the spreadsheet with Claude Code through the Model Context Protocol (MCP) turned what began as a modest macro-repair helper into a full macro-authoring assistant. The shift matters for the large population of office workers who depend on VBA automation yet cannot write it themselves, because it suggests the boundary of what non-developers can build is moving considerably.

The described setup is deliberately simple. The author's monitor is split down the middle: Excel occupies the left half, and Claude Code, Anthropic's terminal-based coding agent, sits on the right. The workbook stays open throughout. Rather than switching contexts or copying code back and forth, the author selects cells, adjusts formatting, and describes what they want in plain Japanese, while the AI on the right responds and operates on the live document.

The starting point was narrower. The author recounts an earlier experience with an [AI Fix] button attached to a macro list, a feature intended to repair broken code with minimal interaction. The headline reframing, captured in the original Japanese title, is that "just talking fixes macros" turned out to be "just talking creates impressive macros." In other words, the conversational interface that was expected to serve as a debugging aid appears to work equally well for generating sophisticated new macros from scratch.

What makes the difference is MCP itself. The Model Context Protocol is an open standard, introduced by Anthropic in late 2024, for connecting large language models to external tools and data sources through a consistent interface. An MCP server acts as a bridge: it exposes specific capabilities, such as reading a worksheet, inspecting a selected range, writing values, or running a macro, so the model can call them as structured actions rather than merely emitting text. Connecting Excel to Claude Code in this way means the AI is not just suggesting VBA code in the abstract; it can observe the actual workbook state and act on it, which tightens the loop between a spoken request and a working result.

That live connection is likely what elevates the experience beyond conventional code assistance. When the model can see the current sheet, the named ranges, and the outcome of running a macro, it can iterate: propose code, execute it, check the effect, and adjust. For someone who understands what they want the spreadsheet to do but not how to express it in VBA, this turns macro development into a dialogue rather than a programming task.

The account fits a broader industry pattern. Since MCP's release, a growing ecosystem of community and vendor-built servers has emerged for tools ranging from databases and file systems to design and productivity applications, and Anthropic has positioned the protocol as a way to standardize how agents reach real-world systems. Claude Code is part of that push, aimed at letting the model take multi-step actions in a developer's environment. Applying the same plumbing to Excel is a natural extension, given how much business logic still lives in spreadsheets and VBA. It also echoes the wider "citizen developer" trend, in which low-code and AI tools let non-programmers assemble automation that once required specialist help.

Some prerequisites and caveats are worth keeping in mind. VBA remains Excel's built-in automation language, and macros carry known security considerations; files with macros are often restricted by default in managed environments, and code that an AI generates should be reviewed before it touches important data. Granting an MCP server write access to a live workbook also means trusting it with real changes, so backups and version control remain sensible. The article is a first-person blog post rather than a controlled evaluation, so its enthusiasm reflects one author's workflow rather than a benchmarked comparison.

Even with those qualifications, the piece is a concrete illustration of how MCP is intended to work in practice: a familiar application, an agentic assistant, and a shared protocol that lets them cooperate on the same document. For readers weighing whether conversational AI can meaningfully accelerate everyday Office automation, the takeaway is that the technology appears to have moved from fixing what exists toward helping build what does not yet exist, at least for the kinds of macro tasks described here.