AI second brain: notes your assistant can actually use

What makes a second brain “AI”

Every notes app now has an AI feature, and the phrase “AI second brain” hides the distinction that matters. Built-in AI means the app summarizes, tags or chats inside its own walls: useful, but it only works while you are in that app, and it knows nothing you didn’t put there. An open interface means an assistant you already use connects to your notes through MCP — the Model Context Protocol, the open standard Claude, ChatGPT, Cursor and most agents speak — and can search, read and, if you allow it, write them from wherever you happen to be working. Only the second makes your notes part of the assistant’s memory rather than a feature of the app: the same library answers in a Claude chat, in your code editor and in an agent running on a server, and it keeps answering when you switch assistants.

That is the whole idea of a second brain AI setup: the notes stay a place you can read, correct and export, and the assistant gets a door with a lock you control. What it is not is an assistant that quietly remembers everything you ever said; that is a different product with a different trade-off, and the OpenClaw memory plugin page explains where Dexi stands on it.

How to build an AI second brain in four steps

1. Pick where captured knowledge lands. The assistant can only search what you saved, so the first decision is the capture path, not the model. In general: one app, with a browser clipper, a phone share sheet, an email address you can forward to and feeds that arrive on their own. In Dexi, every one of those is built in — pages and highlights from the web clipper or bookmarklet, newsletters forwarded to your personal save address, RSS subscriptions, the iOS share sheet — and each arrives as a note, tagged on the way in, next to the notes you type (the capture hub has all of them).
2. Connect the assistant. Add the notes app’s MCP server to the assistant and sign in. In general this is a URL and an OAuth approval for hosted servers, or a local process you run for community servers. For Dexi it is one URL, https://mcp.dexi.net/mcp, added as a connector in Claude, ChatGPT, Cursor or any MCP client, followed by a sign-in on Dexi’s own page — no API key to paste.
3. Set what it may read and write. Decide the boundary before the first question. In general: grant read-only unless you want the assistant filing things, and limit it to the part of your notes it needs. In Dexi the approval page asks exactly this: everything, one folder, one tag, or a tag within a folder, and whether writes are allowed at all. A restricted assistant sees only notes inside the boundary and files new ones there; you can change or revoke it later in Settings → Connected apps, and the change applies on its next request.
4. Use it: ask, file back, review. Ask from your notes (“what have I saved about usage-based pricing, and where do the sources disagree?”), save conclusions back as notes when they are worth keeping, and let a weekly rhythm resurface what you filed. In Dexi the assistant can search by keyword or meaning, read notes in full, list tags and folders, query your saved bases, create and update notes, and run your spaced-repetition review deck in conversation.

How to set up a Claude second brain

Claude is the assistant most people mean when they say “second brain Claude”, and it supports remote MCP servers as custom connectors on its paid plans. The setup is the same on the web and in Claude Desktop: open Settings → Connectors, add a custom connector with the server URL https://mcp.dexi.net/mcp, click Connect, and approve on Dexi’s page — choosing the boundary there. From then on any conversation can draw on your notes, and Claude’s tools show up under the search-and-tools menu in the chat. The click-by-click version, with the plan requirements and troubleshooting, is Connect Dexi to Claude; Claude Code has its own route (a plugin or claude mcp add) at Connect Dexi to Claude Code.

What Claude already holds: its memory feature saves topics about your work and preferences as you chat, and Projects keep their own context. What it holds through Dexi: the thing you read in March, the highlight you clipped last week, the note you wrote after a call — searchable by meaning, so “that piece about why cities scale and companies don’t” finds the note even when none of those words are in it. The two are complementary, and the Claude and ChatGPT with your notes article walks through a research session end to end.

ChatGPT, Cursor and other assistants

The same server works everywhere MCP does. ChatGPT connects through developer-mode connectors on its paid plans — Connect Dexi to ChatGPT. Cursor takes the URL in its MCP settings so the Agent can consult your notes while it codes — Connect Dexi to Cursor. VS Code, Goose, Hermes Agent, OpenClaw and Smithery each have a guide on the MCP overview, and any other client that speaks streamable HTTP with OAuth connects the same way.

Second brain + Obsidian and AI: what works and what doesn’t

The most-watched setup videos pair Claude with an Obsidian vault, and it works: Obsidian has no official MCP server (as of November 2026 it ships a CLI, not an MCP endpoint), but the community has several — the Local REST API plugin now serves an MCP endpoint from inside Obsidian, the MCP Connector plugin does the same, and file-based servers read the vault straight from disk with no plugin at all. Logseq has the same shape through its local HTTP API. The strengths are real: your notes stay local Markdown files, it costs nothing, and nothing leaves your machine except what the assistant asks for.

The costs are real too. You run the server, and keep it running. The assistant’s access is all-or-nothing — the whole vault — unless the particular server adds scoping, and most don’t. A local server is reachable only by clients on that machine, so claude.ai in a browser, a phone, or an agent on a VPS can’t use it without a tunnel. And nothing works when your laptop is closed. If your vault must stay local, this is still the right choice; the trade is upkeep and reach for ownership. If the reason you want an AI second brain is to ask questions from anywhere and have things you clip on your phone answerable the same evening, a hosted connection fits better — which is the whole case for Dexi, and the reason it is a web app rather than a folder of files.

Best AI second brain apps, by how your assistant reaches them

Where each app stands as of November 3, 2026, read on each vendor’s own page. “Official” means the vendor ships and documents the server; “community” means a third-party server you install yourself. Prices are the plan that unlocks the connection, annual-billing rates where applicable. (Disclosure: Dexi is our product. The row is as honest as the others.)

As of November 3, 2026 — Anytype’s storage-tier prices as last read in August 2026 (its pricing page rendered no prices on the re-checks); Tana Outliner’s hosted server is labelled beta on its own page. Check each product’s site for current details.

Three things the table makes visible. First, the field moved: Notion, Heptabase and RemNote all shipped official servers in 2025–26, so “built-in AI only” is no longer the dividing line. Second, hosted versus local now matters more than official versus community — a local server (Anytype, RemNote, Tana’s free tier, every Obsidian option) needs the desktop app open and reaches only clients on that machine. Third, almost nobody scopes: a connection is the whole vault, space or workspace, with Heptabase and RemNote offering a read-only mode and Dexi the only one on this list that limits a connection to a folder or a tag. Choose on those three axes and the shortlist is short.

What an AI second brain can’t do (yet)

It doesn’t remember your conversations unless you file them. An MCP connection gives the assistant tools, not a diary. A conclusion you reach in a chat is gone with the chat unless you say “save this as a note” — and that is by design, because the alternative is a memory full of things you never chose to keep. Agent plugins that write a digest note per session exist (the Dexi plugins for OpenClaw, Hermes and Claude Code offer one, opt-in); the plain connection does not.

It can write the wrong thing. An assistant that can create and edit notes can also create a duplicate, overwrite a line, or file a confident mistake. Dexi’s per-connection boundary limits where it can read and file — it is not a review step. So grant write access to the assistants you would let edit a shared document, keep experimental agents read-only or inside one tag, and treat anything an assistant saved as a draft until you have looked at it.

Retrieval is only as good as what you saved. The assistant searches your notes; it does not read the web on your behalf and back-fill the library. If the capture paths in step one are not wired, the second brain starves and the assistant answers from the open internet instead — which is the failure mode this whole setup exists to avoid.

Frequently asked questions

Keep reading