Dify × Notion MCPで研究室の議事録AIチャットボットを作ろうとして詰まった話 Dify × Notion MCPで研究室の議事録AIチャットボットを作ろうとして詰まった話

Building an internal chatbot that can recall what happened in past meetings is an increasingly common goal for small teams, and a recent blog post documents one student's attempt to do exactly that for a university research lab. The author, who describes themselves as a student studying to become a web engineer, set out to combine Dify with a Notion MCP server so that lab members could ask a chatbot about meeting minutes and assigned tasks rather than scrolling through pages of notes. The post is framed as a record of where the project got stuck, which makes it a useful case study in the practical friction of wiring these tools together.

The motivation is familiar to anyone who keeps team knowledge in a shared workspace. Research labs typically log meeting content, decisions, and individual responsibilities somewhere central, and Notion is a popular choice for that kind of structured note-taking. Over time, however, the volume of pages makes it hard to answer simple questions like who is responsible for a task or what was decided in a particular session. A natural-language interface that reads directly from those notes promises to turn a growing archive into something queryable.

Two pieces of technology sit at the center of the attempt. Dify is an open-source platform for building large language model applications, offering a visual interface for assembling chatbots, agents, and workflows without writing everything from scratch. It handles prompt orchestration, model selection, and increasingly the integration of external tools. The Model Context Protocol, or MCP, is the second piece. Introduced by Anthropic and since adopted across a range of AI clients, MCP is an open standard that defines how an AI application connects to external data sources and tools through a common server interface. A Notion MCP server, in this design, exposes Notion content so that the model can search and retrieve pages on demand rather than relying only on what was baked into a prompt.

In principle, the combination is appealing because it separates concerns cleanly. Dify supplies the conversational front end and the model, while the MCP server handles authenticated access to Notion's data. That modularity is part of why MCP has gained momentum: rather than building a bespoke integration for every tool, developers can reuse a standardized connector. The approach mirrors a broader industry shift toward giving language models structured, permissioned access to live systems instead of static documents.

The reality, according to the post, proved less smooth. While the summary does not enumerate every obstacle, projects of this shape commonly run into a handful of recurring issues, and the author's framing suggests the setup did not work as expected on the first pass. Connecting an MCP server to Dify can require matching protocol versions and transport methods, and support for MCP in application platforms has been evolving quickly, so configuration steps that work in one release may differ in another. Authentication is another frequent sticking point, since reaching Notion's API generally involves creating an integration, issuing a token, and explicitly sharing the relevant pages or databases with that integration. If those permissions are not granted, the server can appear connected while returning no usable content.

Readers considering a similar build should keep several adjacent details in mind. Notion offers its own official API and, more recently, MCP-related tooling, so it is worth confirming which server implementation is being used and how current it is. Dify is not the only option for the orchestration layer; alternatives such as n8n, LangChain, or directly using an MCP-capable client like Claude Desktop can achieve comparable results with different trade-offs. It is also reasonable to expect that retrieval quality depends heavily on how the Notion workspace is structured, since well-tagged, consistently formatted minutes are far easier for a model to search than free-form notes.

The value of a post like this lies less in a finished product than in the documented friction. As MCP continues to mature and more platforms add native support, the rough edges described here are likely to smooth over. For now, the account serves as a reminder that combining a low-code AI platform with an emerging interoperability standard remains an experimental exercise, and that the gap between a promising architecture and a working chatbot often comes down to configuration, permissions, and version compatibility.