MCP: your workspace, inside your AI tool.
secondbrain exposes a remote MCP server. Connecting it to Claude, to ChatGPT or to any client that speaks the protocol gives it access to your notes, tasks, documents, recaps and the whole Atlas — to query, to analyze and to write. This is the channel for the things that do not fit in a chat: long-range planning, cross-cutting analysis and automation.
What MCP is, in one line
MCP (Model Context Protocol) is the standard that lets an AI tool use external services as its own tools. Instead of copying and pasting context into a prompt, the model asks for the data it needs, when it needs it.
The practical difference is enormous: an analysis over two hundred documents does not fit in a prompt, but it does fit in twenty tool calls the model decides to make on its own.
Connecting a client
The server is remote — nothing to install — and answers at a single address:
https://api.secondbrain.media/functions/v1/mcpThere are two ways to authenticate, and you pick by client:
OAuth
The client opens the secondbrain sign-in screen, you authorize, done. This is what Claude and ChatGPT use when adding a remote connector: no key to copy anywhere.
API token
For clients that expect a fixed credential: generate a token in Account settings → MCP, paste it into the client, revoke it whenever you want.
Clients it works with
Any client that supports remote MCP servers over HTTP: Claude (desktop, web and Claude Code), ChatGPT with connectors, editors like Cursor or Windsurf, and your own agents built with the protocol SDKs. The server is also published in the public MCP registry, so clients that consult it find it on their own.
The tool catalogue
Seventy-four tools, grouped by family. A connected client sees all of them and picks the one it needs.
| Family | What it can do |
|---|---|
| Daily notes | Read a day's note, write it, list recent ones, search across all, close and reopen the day |
| Tasks | List, read detail, create, update, complete, roll back, delete, list the completed ones and read the working log |
| Documents | List the tree, read, create, update, move, delete, search, and link or unlink entities |
| Recaps | List meetings and groups, read a whole one, create, update, delete, and manage meeting types |
| Atlas | Search entities, read an entry, create, update, delete, merge, manage relations and tags, read definitions and an entity's tasks |
| Canvases | List, read, create, update and delete |
| Files | List everything, list by document, rename and delete |
| Reports | List, read the latest or any one, generate a new one, delete, and check AI usage |
| Workspace | List your workspaces, read and update the current one, and read its relations |
| Account | Read and update the profile, check the plan and storage usage |
| Search | Hybrid search across the whole workspace in a single call |
What it is actually used for
Quick capture is already handled by mobile and chat. MCP is for the rest: what needs to read a lot and think for a while.
Complex planning
"Look at my open deliverables, my meetings for the next two weeks and my real load, and build me a quarter plan." The model reads, cross-references and proposes; you decide what gets written.
Atlas analysis
"Which clients have stalled projects, and who on my team is connected to them?" It walks the graph, reads the entries and answers from what is there, not from an impression.
Synthesis over lots of material
Reports that cross months of recaps, notes and documents. The model asks for what it needs across several rounds instead of trying to cram everything into one prompt.
Automation
Routines that run on their own: prepare Monday's summary, open tasks from an email, archive what closed, create the entries for a new client.
Bulk loading and migration
Bring content over from another tool, create hundreds of documents with their structure, or register an entire Atlas from a list.
Writing with your context
Draft in Claude or ChatGPT with your documents and recaps as the source, and save the result straight into your memory.
What you control
- A token grants access to your account and your workspaces, and nothing else. The same boundary the app has.
- Each token is revoked separately from Settings, without touching the others or ending your session.
- Tools respect your plan: what is not included cannot be done over MCP either.
- Every write is logged exactly as if you had done it in the app, with its history and its rollback.
- You can disconnect the server from the client at any time: nothing keeps running on our side.
Combining it with other tools
The point of MCP is that your AI client can have several servers connected at once, and use them in the same conversation.
With secondbrain plus your calendar, your mail, your repository or your database connected, a single request can read your work context here, cross it with what is over there and leave the result written where it belongs. secondbrain contributes the part none of the others have: what you were working on, with whom and why.
Generate an API token →Integrations
MCP
External AI clients (Claude, ChatGPT and others) can query and operate your workspace over the MCP protocol, only with the token you generate.
Shall we try it?
Create your free account and use all of this today.