「Notionで良くない？」と思ってた私がObsidianにハマって、ついにClaudeと連携させた話 「Notionで良くない？」と思ってた私がObsidianにハマって、ついにClaudeと連携させた話

Many engineers have heard of Obsidian, but plenty initially dismiss it with a shrug and the thought, "Isn't Notion good enough?" That was the starting point for the author of this piece, who eventually changed their mind, embraced Obsidian as a daily note-taking tool, and went further by connecting it to Claude through the Model Context Protocol (MCP). The shift matters because it reflects a broader pattern: developers increasingly want their personal knowledge bases to be readable not just by humans, but by large language models that can search, summarize, and reason over them.

The skepticism is understandable. Notion offers polished collaboration, databases, and a hosted experience that works out of the box. Obsidian takes a different approach. It stores notes as plain Markdown files on local disk, with linking between notes forming a personal "graph." There is no proprietary database; your vault is just a folder. For people coming from Notion, the early impression can be underwhelming, since the interface looks plain and many features arrive only through community plugins. But that local-first, file-based design is precisely what makes the later integration with an AI assistant feasible. Plain text is easy for tools to parse, version with Git, and feed into a model.

The bridge in this story is MCP, an open standard introduced by Anthropic to let AI assistants like Claude connect to external tools and data sources in a consistent way. Rather than building a custom integration for every app, MCP defines a server-client model: an MCP server exposes resources and actions, and a client such as Claude Desktop calls them. Several community MCP servers exist for Obsidian, typically connecting either to the local vault directly or through the Local REST API plugin. Once configured, the assistant can list notes, read their contents, search by keyword, and in some setups create or update files. This turns a static archive of Markdown into something the model can actively work with.

In practice, the setup usually involves installing an MCP server, pointing it at the vault path, and registering it in Claude's configuration file. From there, a user can ask the assistant to find related notes, draft a summary across several documents, or restructure scattered ideas into an outline, all while the underlying files remain plain text that the human still owns. The appeal is less about novelty than about reducing friction. Knowledge that was previously trapped in separate apps becomes queryable, and the assistant operates on real personal context rather than generic web knowledge.

Some background helps frame why this feels significant now. The original summary suggests the writer's surprise that Obsidian, despite its rough first impression, proved more flexible than expected once they invested in it. That mirrors a wider industry move toward local-first, AI-augmented workflows. Tools like Logseq and the Markdown-based notes in many editors share the same philosophy, and even Notion has added AI features and exposed APIs. MCP itself has gained traction beyond Anthropic; the protocol is open, and a growing number of clients and servers cover services from file systems to databases. That ecosystem momentum is likely part of why pairing Obsidian with Claude appears increasingly common among developers.

There are caveats worth keeping in mind. Granting an assistant access to a personal vault means thinking about which notes it can read and whether it can write, since accidental edits or data exposure are real risks. Community MCP servers vary in maturity, and configurations can differ across operating systems, so the smoothness of the experience is likely to depend on the specific tools chosen. None of this is a turnkey consumer product yet; it remains the kind of integration that rewards comfort with config files and command lines.

The takeaway is modest but practical. Notion and Obsidian are not strictly competitors so much as different bets, one on hosted convenience, the other on local control and extensibility. For someone who values owning their data and wants an AI assistant grounded in their own notes, the Obsidian-plus-MCP path offers a credible option. The author's arc, from doubt to enthusiasm, suggests the value emerges only after the initial setup pays off, which is a fair reminder that these workflows still ask for patience before they feel indispensable.