Obsidian MCP: ways to give an AI agent your vault

Pick by two questions: does the agent run on your own computer, and does anyone besides you need the same vault? For one person, a file-based MCP server such as MCPVault, or Claude Code opened in the vault folder, works with Obsidian closed. Servers built on the Local REST API plugin need Obsidian open. None of them gives a team roles or shared editing; for that, put the vault in Git and serve it from a team wiki such as Folio.

Below: six ways, one table with the same five criteria for each, and what each is good for. Facts about other projects come from their own pages and were read on 2 November 2026; they change, so check the linked source before you rely on one. If you already know you want the team route, the team route is the last one in the table, and the guide to making a vault multiplayer has its setup.

Which way of giving an agent your Obsidian vault fits you?

A Git repository with an agent cloned next to it is a seventh route, and it sits between the folder and the team routes: the rights are whatever your Git host grants, and several people can work because Git merges their commits, though not live. Folio builds on exactly that.

How does the Local REST API plugin give an agent access?

The community plugin Local REST API serves your vault over HTTPS on 127.0.0.1, port 27124 by default, with a key from its settings. Its README says it also contains an MCP server, so a client such as Claude can connect without extra software. It runs inside Obsidian, which means the app has to be open, and the key is one secret for the whole plugin.

The older route is mcp-obsidian, a separate Python server that talks to the plugin. Its tools list and read files, search, patch or append content, and delete files. It needs the plugin installed, the key in an environment variable and Python 3.11 or later. Both are MIT-licensed. Pick this family when you want the agent to use what Obsidian already knows, at the price of keeping Obsidian running.

What changes with a file-based MCP server?

A file-based server skips Obsidian and reads the Markdown files. MCPVault describes itself as a local server with 18 tools for files, search, front matter, tags and wiki links. The client starts it as a local process and passes the vault path. It keeps to the vault root, refuses dotfolders such as .obsidian and .git, and has a read-only flag. Nothing needs to be open, and nothing leaves your machine except what the model provider receives.

Start read-only. An agent that can write to your notes can also damage them, and a vault that is not in Git has no history to roll back to.

Can an agent just work in the vault folder?

Yes, and for one person it is often the simplest answer. A vault is a folder of Markdown files. Open Claude Code or another coding agent in that folder and it reads, searches and edits the notes with its ordinary file tools. Put the rules of the vault in a CLAUDE.md file so every session starts with them; Anthropic describes how these files and auto memory work in the Claude Code memory docs. The practical walk-through is in Obsidian and Claude Code: a shared second brain for a team.

The weak spot is the same as with every local route: the vault lives on one disk, so only the person at that computer has the agent's view of it.

Is the Obsidian CLI an option?

Obsidian ships a command line interface that needs the 1.12.7 installer, is enabled in Settings, and, as its page says, requires the app to be running. An agent with a shell can call it. It is not an MCP server, so a client that only speaks MCP cannot use it directly.

How do you give a whole team's agents the same vault?

Move the single copy off one laptop. Put the vault in a Git repository, connect the repository to a server every colleague can reach, and give the agent a token for that server instead of a path on a disk. Folio (foliowiki), an open-source (MIT) self-hosted team wiki in Git, does this: it opens an Obsidian vault in the browser without changing a file (see Folio vs. Obsidian), and agents reach the same pages over MCP with the rights of whoever owns the token.

To connect Claude Code, create a token under the user menu, API tokens, and run one command. Choose the read scope while you are testing.

claude mcp add --transport http folio https://your-folio.example.com/mcp \
  --header "Authorization: Bearer folio_pat_..."

Claude Code's own page on connecting to MCP servers covers the options, and Connect Claude Code to Folio has the shared project file and the OAuth route. Other clients (Claude, ChatGPT, Cursor, Codex, VS Code) have pages under integrations.

What the agent gets beyond reading files is the structure of the vault. Search runs across the space, and the tools return backlinks with the line each link sits on, a page's outline and properties, the link graph, unresolved links and orphan pages, and the rows of an Obsidian base. An agent can ask the vault what is broken, not only what is written. Everything is checked against the token owner's rights, so a page that person cannot open does not appear, and a read token cannot write. Writes come back as ordinary Git commits.

What does Folio not do here?

• It is not a local server for a folder. Folio clones a Git repository onto its own server. It cannot be pointed at a vault that lives only on your disk, so put the vault in Git first.
• It does not run Obsidian plugins. Dataview and query blocks are shown as code, community plugins and their CSS do not exist in the browser, and Kanban and map views of a base fall back to a table.
• It reads Obsidian markup and never writes it. It does not turn your links into wikilinks. A page an agent creates through MCP is plain Markdown.
• MCP does not manage rights. Page access, roles and the creation or deletion of a space are done in the interface. An agent cannot widen its own access.
• Sync is not instant. A browser or agent edit becomes a commit after about 90 seconds of quiet, and Folio fetches from the remote about every three minutes.
• You run the server. It needs PostgreSQL and Redis next to it. For one person on one laptop, a file-based server is less work.

The complete list is in LIMITATIONS. To see the screens first, the public demo runs the release; the install guide has ten ways to run your own.

All posts