[DigitalToday reporter Seulgi Son (손슬기)] A local note app, Obsidian, is spreading quickly among developers as an AI coding assistant tool. They organise requirements and design intent for AI in Obsidian and use it as a knowledge base.

If the Model Context Protocol (MCP) is connected to Obsidian, AI automatically references documents and generates code. The reporter also tried Obsidian.

Obsidian is a note app that stores markdown files locally. What sets it apart from typical note apps is links between notes. If a specific word or document title is wrapped in square brackets ([[, ]]), it creates links between notes. These relationships are visualised as a graph. If an "Interview A" note links to "Source B", you can see at a glance which articles and memos Source B connects to. Unlike traditional note apps that stack files in folders, it is structured to accumulate relationships between memos.

Obsidian is drawing attention in AI coding because it is advantageous for context management. To get good output from AI, you need to select and provide only the necessary information. There is an input limit in the context window, and output quality declines if too much information is added at once. If requirements, design intent and data flow are organised into notes, AI can select and read only the needed documents and generate more precise code. Adding MCP automates this process. If you connect an Obsidian MCP server to Claude Code, AI opens documents in the vault, follows backlinks to related notes, grasps the context and generates code. The step where people copied documents and entered commands disappears.

The reporter tested Obsidian using a scenario of building an AI companion app. It was necessary to check whether the conversation response generation module worked properly, so a test harness was designed. A test harness is a verification framework that automatically checks whether specific functions work as intended without touching production service code. Obsidian was used for this.

The first hurdle was designing the folder structure. It was necessary to decide in advance what to build, in what structure and in what flow, but it was hard to get a feel for it.

With Claude's help, it was divided into 4 folders and 13 notes. Requirements (01_requirements) stored functional requirements and constraints, architecture (02_architecture) stored component structure, data flow and file layout, prompts (03_prompts) stored drafts of instructions to pass to AI, and results (04_results) stored experiment results. Each note listed the role of each component, such as test execution, data injection and substituting external APIs, and was connected through backlinks. Reference paths were designed in advance so AI could naturally move from one note to related notes.

Once the design documents are in place, turn on Claude Code, install the Obsidian MCP server package and register the vault path. After that, you just ask it to write the harness code by referencing the markdown file names. AI opened the relevant notes, followed backlinks, explored related notes and generated code. There was no copy-and-paste and prompt-assembly process. When the generated code was run with pytest, all 6 test cases passed. Writing a verification environment of this level from scratch could take half a day.

CODE STRUCTURE NEEDED... MCP STABILISATION ALSO A TASK

From a non-developer's standpoint, Obsidian has a barrier to entry. Deciding how to divide the folder structure and what information to include in notes requires an understanding of code structure. Even if your usual memos have accumulated, you still need a separate process to change folder names or reclassify notes.

The MCP connection process was not smooth either. Response delays of hours repeatedly occurred while the Obsidian MCP server was exploring the vault. Many similar error issues have been posted on the GitHub developer community as well. The MCP ecosystem itself is still at an early stage. Improving stability remains a task.

It was rather inefficient for one-off or simple tasks. In terms of speed, it is much faster to handle it directly in Claude Code or inject context into a Claude project and generate artifacts.

For that reason, Obsidian is assessed as showing its true value in complex, long-term projects where multiple modules are intertwined and requirements change frequently. The more design documents accumulate, the richer the context AI can reference becomes. Instead of explaining from scratch every time a new feature is added, you only need to add notes to the vault.