Notion MCP: Connecting a Knowledge Base to a Local AI
Connecting the official Notion MCP to Doka through OAuth: searching a knowledge base, building reports, and updating pages without pasting notes into a chat.
A knowledge base is only useful as long as you can actually find things in it. Once it grows to a few hundred pages, "where did we write down the decision on releases?" turns back into a question for the work chat. Notion MCP lets you ask an agent instead: it finds the right pages, pulls the facts together, and, if you allow it, updates the document.
What you can do through Notion MCP
The official Notion server can search a workspace, read and create pages, update tasks, and pull reports together from several sources. The scenarios are pretty down to earth:
- gather decisions on a project from meeting notes;
- turn research notes into a draft technical spec;
- prepare a weekly status from a task database;
- find outdated instructions and suggest a new version.
The agent only gets the access you approve through OAuth. It's still worth starting with a small team or a separate workspace, rather than opening up the entire company knowledge base right away.
How to connect Notion MCP in Doka
The official Notion MCP works
as a remote server. Doka runs MCP over local stdio, so you need Notion's recommended
mcp-remote bridge in between.
In Doka, open "MCP Servers" → "+ Add server" and set:
- Name:
Notion. - Command:
npx. - Arguments, one per line:
-y,mcp-remote,https://mcp.notion.com/mcp. - Environment variables: leave empty.
On the first request, a browser window opens: log into Notion, pick a workspace, and approve access. There's no token to copy by hand. After authorizing, ask the agent to find a specific page — that's the simplest test.
Notion MCP can do more than read — it can change pages too. For your first tasks, phrase it as "prepare a draft, don't write anything yet," and confirm changes separately.
Is this local work or cloud work
Worth not blurring two things here. The agent and the model can run on your own computer, so page content doesn't get sent to an outside model provider. But Notion is a cloud service, and the official MCP server is hosted by Notion. Requests to the knowledge base go through their infrastructure.
If you need a fully isolated setup, Notion as a source doesn't fit that anymore — in that case it's better to keep documents inside your own network and connect local storage instead. An older open-source Notion server with an API token exists, but Notion itself no longer treats it as the primary, actively maintained option.
How to avoid a mess instead of an answer
Don't ask it to "go through all of Notion." Give it a scope and a freshness cutoff: "use pages about project X updated after June 1, and attach links to sources." For a report, define the structure up front: decisions, open questions, owners, deadlines.
If the agent is going to update a page, ask it to describe the diff in words first — what it'll add, remove, and why. That costs an extra minute, but it protects a live knowledge base from a confident but wrong edit.
Other integrations are collected in the list of useful MCP servers, and the general field setup is covered in the step-by-step guide.
And a caveat on boundaries, since we're talking about a knowledge base. Notion is a cloud service, and its MCP server reaches it over the network — a local agent doesn't change that. Everything else stays local: the agent itself, your files, and the model, if you picked a local one. That's a reasonable trade-off when your notes already live in Notion, but it's worth knowing upfront. Download Doka for free.