An MCP second brain is only as good as where its notes live

An MCP second brain is one store of notes with a Model Context Protocol server in front of it, so any assistant that speaks MCP, Claude, ChatGPT, Cursor or Claude Code, can search, read and write the same notes. You write a decision down once and every tool you use can find it, instead of re-explaining it in each new chat.

The idea is simple. The choices that matter come after it: where the notes physically live, what format they are in, and what happens when more than one person, or more than one assistant, writes to them. This page compares the real options on those three questions, checked against each product's own pages on 13 October 2026.

What makes an MCP second brain work?

Three properties, in order of how often they get skipped.

The notes have to outlive the tool. If the store is a proprietary database, your second brain is only as durable as that company. Plain markdown files in a folder are readable by anything, forever, and an assistant writes them the way a person would.

Writes have to be safe. One person with one assistant can get away with an assistant rewriting a whole file. Two assistants, or an assistant and a teammate, cannot: the slower writer silently undoes the faster one. A good MCP server edits by operation and refuses a write based on a stale copy.

Access has to follow the person. The moment the notes are shared, the server has to know who is asking. An assistant connected with your key should see what you can see, and nothing in the folder you were never given.

What are the options for an MCP second brain?

Obsidian plus an MCP server is the most common setup. The servers are community projects, most of them MIT licensed, and they either talk to Obsidian through the Local REST API plugin, which means Obsidian has to be open, or read the vault folder directly. Each runs on one machine with one person's access. We compared the routes in three ways to run an MCP server for Obsidian.

Basic Memory is the closest in spirit to a plain-files second brain: the local version is open source under AGPL-3.0 and writes markdown files on your disk. Its cloud Team plan is $15 per seat a month, falling to $12 at 20 or more seats, and lists "Private, exportable Markdown files".

Hjarni is a hosted markdown knowledge base with the MCP server built in. It is free for 25 notes, Pro is $12 a month, and Teams is $20 per seat a month or $200 per seat a year. On self-hosting its FAQ is direct: "Not currently. It's a hosted service."

Notion's MCP server is hosted by Notion and reads and writes pages and databases. Notion's own guidance is to "Start with single-player use cases", because the server acts with the user's full Notion permissions.

Should the notes live on your disk or in the cloud?

Both work with MCP. The difference shows up later.

A hosted store is easier on day one: one URL, OAuth, no server to run, and cloud assistants like ChatGPT can reach it directly. The cost is that the notes are a copy in somebody else's system, and leaving means an export.

A store on disk is yours outright, and a local agent like Claude Code can work on the files with no server at all. The catch is reach. A cloud assistant cannot connect to localhost on your laptop, so to use ChatGPT or Claude on the web against local notes, something has to expose a public HTTPS endpoint.

The useful middle is notes that live on every person's disk as files and also sync through a server that speaks MCP. That is the shape Baalda takes.

What changes when a team shares the second brain?

Everything that was fine for one person becomes a bug.

• Keys. A single-user MCP server holds one key over the whole vault. Share it and every teammate's assistant has the owner's access to everything.
• Concurrent writes. Two assistants appending to the same meeting note at once will overwrite each other unless the server merges by operation.
• Attribution. "The AI changed it" is not an answer when three people's assistants write to the same notes. Each write has to belong to a person.

We wrote up the access side in detail in an MCP server for team notes has to know who is asking.

How does Baalda handle MCP?

Baalda serves MCP at /api/mcp on its server, over Streamable HTTP. On the managed service that is https://api.baalda.com/api/mcp, and on a self-hosted server it is your own URL with the same path.

• Per person, per vault. Each token is minted by one person for one vault in Vault Settings, under MCP, and carries that person's access. Claude and ChatGPT can also connect through OAuth, choosing the vault on the consent screen.
• The same permissions as the editor. A folder set to No access for you is invisible to your assistant too, because the server checks the same per-file rules the app does.
• Edits merge. Writes go into the same CRDT the editors use, as changes to the text rather than whole-file replacements, and a write can carry the revision it was based on so a stale one is refused instead of applied.
• Files on disk. What an assistant writes becomes an ordinary .md file on every member's machine, readable without Baalda.

Where it is the wrong pick: if you are one person who wants a second brain for Claude Code only, a plain folder needs nothing else. And if your notes are already in Notion and the team is happy there, Notion's own server is the shorter path. For the one-person case in more depth, see the best second brain for Claude.