VBAマネージャーの開発から1年で、Excelの中に住むエージェントの初号機を作った話One year after starting an AI-driven Excel VBA manager, the author built a…

An independent developer has marked the first anniversary of an ongoing effort to drive Microsoft Excel's VBA macros with AI, publishing an account of a "first-generation" agent that runs inside the spreadsheet itself. The project, described in a Qiita blog post, treats Excel less as a passive calculation grid and more as an execution environment in which an autonomous agent can operate — a framing the author connects to the Model Context Protocol (MCP).

According to the post, the author began building "a tool to operate Excel VBA with AI" in August of the previous year, and the milestone marks exactly twelve months of development. Over that period the project has accumulated a series of capabilities, each documented in separate write-ups. The agent described here is presented as an early prototype rather than a finished product, and the author frames it as one more installment in a continuing series.

MCP is an open standard, introduced by Anthropic in late 2024, that defines how AI applications connect to external tools, data sources, and services. A program that exposes capabilities through the protocol is generally called an MCP server, while the AI-facing application that consumes them acts as the client. The approach has gained traction across the industry because it offers a consistent interface for wiring language models into software they would otherwise be unable to reach directly. The tags accompanying the post — including mcp-server, agent, and ai-automation — suggest the author is packaging Excel's automation surface behind such a server so that a model can call into it.

VBA, or Visual Basic for Applications, is the macro language embedded in Microsoft Office. It has powered spreadsheet automation for decades and remains widely used in finance, operations, and reporting workflows, despite Microsoft's more recent push toward Office Scripts and the Python-in-Excel integration. Bridging VBA to a modern AI agent is notable precisely because so much existing business logic still lives in these macros; an agent that can read, write, and run them could interact with a large installed base of legacy automation rather than requiring everything to be rebuilt.

The author situates the new agent within earlier work. A previous article described a "three-tier" approach to locating files, escalating from a filename match to a full-text search of file contents and finally to an AI-assisted step when the simpler methods fall short. That layered design — trying cheap, deterministic methods before invoking a model — is a common pattern in AI tooling, where model calls carry cost and latency. The author had also signaled that an "AI summary explanation" feature was coming next, but chose to publish this agent piece first.

The broader significance is that the effort sits at the intersection of two trends. Agentic AI, in which models are given tools and allowed to take multi-step actions rather than simply returning text, has been a central theme across the industry through 2025 and into 2026. At the same time, Microsoft has been embedding its own Copilot assistant into Excel and the wider Microsoft 365 suite. An independent, MCP-based agent that lives inside Excel represents a grassroots take on the same idea, built around an open protocol rather than a vendor-specific assistant, and it points to how individual developers are experimenting with the same architecture the major platforms are commercializing.

As a personal project chronicled on a developer blog, the write-up appears to be exploratory, and the "first-generation" label suggests the author expects significant revision. The post does not, based on the available excerpt, claim production-grade reliability or broad compatibility, and readers evaluating the approach would likely need to consult the full article and any accompanying code for implementation specifics. Even so, the account is a useful data point on how MCP, VBA, and agentic patterns are being combined to extend a tool that remains a fixture of everyday office work. Whether such in-application agents prove practical at scale is still an open question, but the year-long progression documented here shows the concept moving from a simple automation helper toward something closer to a resident assistant.