Connecting Obsidian and Claude: three ways compared

An Obsidian vault is a folder full of Markdown files — which is exactly why there isn’t one way to connect it to Claude, but three. All three work. They differ in what you have to put in and what comes out the other end. Here they are, honestly side by side, with the table at the bottom.

Route 1: Share the vault as a folder

The most direct route, and it costs nothing: start Claude Code in the vault folder (or share the folder with Claude Desktop) and drop in a CLAUDE.md that explains the structure. Claude then reads and searches the files like any other project.

What it does well: For text notes on one machine this is hard to beat. No setup, no cost, and data only leaves your machine for the model request itself. You version changes with Git.

Where it ends: The access exists only on that machine — on your phone and in claude.ai the folder doesn’t exist. Claude searches with file tools (grep), not semantically: it finds “notice period” when the text says “notice period”, but not when the note talks about “contract duration”. And a scan in the vault stays a file name.

Route 2: A local MCP server

The community maintains several Obsidian MCP servers (usually started via npx, configured with your vault path). Claude gets real tools — read note, create note, search vault — instead of raw file access.

What it does well: A cleaner interface than route 1, sometimes with better search. Creating and editing notes is structured. Everything stays local.

Where it ends: It remains DIY: install Node, maintain a config file, track updates. The server runs while your machine runs. It answers neither the multi-device question nor the scan question.

Route 3: The hosted knowledge graph

Import the vault into Memexa as a ZIP (wikilinks survive), then connect Claude via MCP — the three-minute guide is here. From then on every client reads the same state: Claude Desktop, claude.ai, Claude Code, and ChatGPT or Cursor too.

What it does well: Hybrid search (semantic + full-text) instead of grep. Documents and scans go through text recognition and become citable — source down to the page. Works on every device without anything running on your machine.

Where it ends: It’s a hosted service — your data lives with a provider (in our case: EU, GDPR, Markdown export any time), and beyond a certain usage it costs money. If you work strictly “nothing leaves my machine”, take route 1 or 2.

The table

	Vault folder	Local MCP server	Hosted graph (Memexa)
Setup effort	Minutes	Node + config upkeep	Minutes (URL + sign-in)
Multiple devices / claude.ai	No	No	Yes
Documents with scans	No	No	Yes (OCR)
Page-level sources	No	No	Yes
Where the data lives	Your machine	Your machine	EU servers
Cost	$0	$0	Free plan, then from €12/mo

The honest reading: on “where the data lives” and on cost, DIY wins — and if all you need is your own text notes on one machine, route 1 finishes the job. The two rows where the hosted route stands alone — scans and sources — only matter if your knowledge includes documents. But then they decide everything.

Frequently asked questions

Do I have to give up Obsidian?

No. The usual setup is a division of labor: Obsidian stays the writing tool, the Memexa import runs periodically, and Claude works on the graph. Wikilinks survive the import as real edges.

What happens to my plugins and templates?

They stay in Obsidian and keep working — the import reads the Markdown files, not the Obsidian configuration. Dataview queries and similar plugin syntax arrive as text, not as executed queries.

Does this work with ChatGPT instead of Claude?

Yes — routes 2 and 3 speak MCP, which ChatGPT supports too. Route 1 is Claude-specific (folder sharing + CLAUDE.md).