Obsidian is a note-taking app, but that description undersells it. The thing that makes it different from Notion, Roam, or any of the cloud-based alternatives is this: your vault is just a folder of Markdown files on your computer. No proprietary database, no sync service you're locked into, no account required. Just .md files sitting on disk, organized however you want.
That simplicity is also what makes it uniquely useful as a context layer for AI.
What Obsidian Actually Is
The core features: a Markdown editor with live preview, a graph view that visualizes how your notes connect via backlinks, and a community plugin ecosystem with over 1,000 plugins. It runs on Mac, Windows, Linux, iOS, and Android. Sync is optional and paid, but your vault works completely offline.
What you build in Obsidian is a personal knowledge graph. Each note links to others via [[wikilinks]]. Over time, the graph view shows you which ideas are densely connected and which are isolated. It's the closest thing to externalizing your thinking that I've found in a tool.
The local-first architecture isn't just a privacy feature. It's why AI integration is so much simpler here than in cloud-based tools.
Three Ways to Wire It Into Claude
Because the vault is a folder of .md files, any agent with filesystem access can read it. Here's how I think about the options, from simplest to most capable.
1. The Filesystem Shortcut
Claude Desktop has a native filesystem MCP connector. You point it at your vault folder and Claude can read and search your notes in any conversation. Three clicks in the settings, no plugins. You can ask Claude to summarize everything you've written about a topic, find contradictions across notes, or draft something using your own writing as context.
The limitation: read-only by default, and search is keyword-based unless you add more tooling.
2. MCP Server with Semantic Search
The obsidian-mcp-tools plugin installs as a community plugin and exposes an MCP server that Claude Desktop, Claude Code, or Cursor can connect to. It runs semantic search entirely on-device via Transformers.js, using all-MiniLM-L6-v2 (about 25 MB, downloaded once and cached). No cloud, no data leaving your machine.
The difference from keyword search is real. "Find my notes about API rate limiting" in keyword mode returns notes with those words. In semantic mode it also returns your notes on backoff strategies, retry logic, and queue-based request handling, because those concepts are related even if the words don't match.
The mcp-obsidian plugin from MarkusPfundstein is the most established option (around 3,000 GitHub stars) and gives Claude read-write access, but it requires Obsidian to be open and running since it routes through the Local REST API community plugin.
3. Claudian: Claude Code Inside Your Vault
Claudian embeds Claude Code directly as a sidebar panel in Obsidian. Your vault becomes Claude's working directory. It can read, write, search, and run commands from inside Obsidian itself. This is the most agentic setup. If you already live in Claude Code, this makes your vault a first-class part of that environment.
The Local LLM Path
If you want nothing leaving your machine at all, the combination of the Smart Connections plugin (vault-wide semantic search using local embeddings via nomic-embed-text) plus the Copilot plugin pointed at an Ollama endpoint gives you a fully local AI assistant over your vault. Llama 3.1 8B or Qwen 2.5 7B work fine for retrieval and simple synthesis. They're noticeably weaker than Claude Sonnet on complex reasoning tasks, but for "find me everything I have on X" use cases the quality is acceptable.
Real Use Cases That Actually Work
Inbox processing. Drop everything into an Inbox folder: voice memo transcriptions, rough notes from meetings, links you want to save. Then tell Claude with MCP write access: "Process my inbox." It creates structured notes with frontmatter and tags, links to related notes, and archives the raw captures. What used to take 20-30 minutes happens in under 5. This is the workflow that made me take the whole setup seriously.
Meeting notes with memory. I write meeting notes with a consistent template. With Claude connected, I can ask "What did we decide about the API schema in Q1?" and get a real answer pulled from the right notes, not a generic response or a blank stare.
Draft with your own context. Feed Claude notes from your vault when drafting documentation or a design doc. The output reflects how you actually think about a problem, not a generic synthesis of the internet. For technical writing where you have years of notes on a specific system, this is genuinely useful.
Architecture decision records for agents. A few people I know use the MCP integration so their Claude Code agent can reference ADRs and project docs while working in the codebase. Ask the agent to check your vault for prior decisions on a topic before proposing a solution. It reduces the "I already thought through this" rework.
Where It Falls Apart
The setup only matters if you actually write notes. A vault with 50 notes is not a knowledge graph, it's a file system with extra steps. The AI layer gets more useful as the vault gets denser. If you're starting from zero, budget a few months before it feels like anything more than a fancy search tool.
Retrieval still isn't perfect. Semantic search beats keyword search, but poorly structured or inconsistently written notes return weak results. The quality of what you get out scales with the quality of what you put in. That's not a limitation of the AI integration, it's just how knowledge systems work.
The read-write MCP setups (anything that lets Claude modify notes) require real care about what you're pointing the agent at. I keep Claude write access scoped to specific folders and don't let it touch anything outside my Inbox and Drafts. Letting an agent loose on your entire vault without guardrails is a bad idea.
My Take
Obsidian is the right foundation for an AI context layer because the vault is just files. That sounds mundane, but it means you can point any agent at it with minimal friction, switch AI providers without rebuilding your knowledge store, and own your data completely.
The MCP approach is where the real payoff is for builders. Not a chat interface bolted onto a database, but an agent with read-write access to a graph of your actual thinking. Once I set that up, I stopped treating Obsidian as a notes app and started treating it as persistent memory for my agent stack.
That shift is worth more than any specific feature in the plugin list.