Confluence MCP, Notion MCP and a wiki that is its own MCP server

Atlassian and Notion both run an official MCP server now, so Claude, ChatGPT, Cursor or Codex can search and edit your Confluence or Notion pages. There is a third route for teams that keep their wiki in Git: the wiki itself is the MCP server, on your own host. This post compares the three by what an agent can do, whose rights it uses and where the content lives, and shows how to connect Folio.

What Atlassian and Notion offer was checked in their documentation on 19 October 2026. Both change fast, so use the links to see the current state.

Confluence MCP: Atlassian's remote server

Atlassian calls it the Rovo MCP server, and it is one endpoint for several products. Its documentation lists Jira, Jira Service Management, Confluence and Bitbucket, among others. Sign-in is OAuth 2.1, with API-token authentication as an option, and actions respect each user's existing access controls and permissions. The page names ChatGPT, Claude, GitHub Copilot CLI and Gemini as supported clients, gives setup for VS Code, Cursor, Claude Code, Claude Desktop, Codex and Windsurf, and says other MCP clients can use the server address. MCP calls use Rovo credits from a pool shared by the organization.

For Confluence, the tool list has three groups. Read covers pages, comments, spaces, version history, attachments and export to PDF or Word. Search uses CQL. Write creates and updates content, adds comments, copies, moves and archives pages, and changes permissions and public links. Because the write group can change who sees a page, Atlassian's own page advises least privilege and a review of high-impact changes.

That is a wide reach, and for a company whose knowledge already lives in Confluence Cloud it is the right answer: no migration, and the agent works as the signed-in user. The pages we read do not say which Confluence deployments the tools cover, so check that before you plan around Data Center.

Notion MCP: hosted by Notion

Notion MCP is a remote MCP server hosted by Notion. A client connects with OAuth and can read and update the content that the signed-in person can access. Workspace owners decide which clients are allowed under Settings, Notion MCP; limiting which clients members may connect needs the Enterprise plan, and organization owners can list and revoke connections through the Admin API.

The tool list covers search, fetching a page or database, creating, updating, moving and duplicating pages, databases and views, comments, and moving items to the trash, which asks for a confirmation flag. Some tools depend on the plan: querying several data sources at once and editing or resolving comments need Business or Enterprise (the first also needs Notion AI), and search is AI search on plans with Notion AI and keyword search otherwise.

What the two have in common

Both are a service run by the vendor, in front of the vendor's database. That is convenient, and it fixes the answers to three questions.

• Where does the content live? In the vendor's cloud, in the vendor's format. To get Markdown or Git history you export and convert it first, as in our post on exporting Confluence to Markdown.
• Who is the agent? The signed-in user. Its reach is that user's reach, which is the right default and the reason to give an agent its own account.
• Who decides the tools? The vendor. You use what it ships, and what it charges for or limits by plan.

A wiki that is its own MCP server

Folio is a self-hosted wiki whose pages are Markdown files in Git. It has an MCP endpoint at /mcp on your own address, with 40 tools: 20 that only read and 20 that write (the full list). What that changes:

• The agent has its owner's rights and no more. A personal access token or an OAuth grant acts as the person who made it, narrowed by its scope, read or write. A page that person cannot open stays invisible to the agent: in search, in the tree and in the link graph.
• Edits are ordinary Git history. A write reaches the page file at once, and the commit follows after about 90 seconds of quiet, authored by the token's owner. An agent can read earlier versions of a page, and a bad edit is reverted like any commit.
• People see what the agent did. The agent works on the same pages as people. Its page update merges with live editing, so a person who has the page open sees the change arrive.
• More than pages. Over the same endpoint an agent works with data tables, forms, whiteboards, Obsidian bases, the link graph, tags and properties.
• Some things stay with people. No tool manages access, invites users, or renames and deletes spaces. A read token cannot write at all. A public space is readable by anyone in the browser, but /mcp still needs a token or an OAuth grant.

One warning applies to every wiki MCP, ours included: the text of a page is data, not instructions. Someone can write a page that tries to steer the agent that reads it. Start with read access, and give an agent its own user with the editor role only in the spaces it needs.

Connect Folio to Claude, ChatGPT, Cursor and Codex

The route depends on where the client runs. claude.ai, Claude Desktop and ChatGPT call your server from the vendor's cloud, so your Folio must be reachable over HTTPS from the internet, and they connect with OAuth. Claude Code, Cursor, VS Code and Codex run on your machine, so a Folio on a private network is enough and a token does the job. Each client has a page with exact steps: Claude, ChatGPT, Claude Code, Cursor and Codex. The hub is Integrations.

Claude Code, with a token

Create a token under the user menu, API tokens, with the read scope to start, then add the server.

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

Cursor and Codex

Both read the token from an environment variable. Cursor takes it in its MCP file, Codex in its config file.

{
  "mcpServers": {
    "folio": {
      "url": "${env:FOLIO_URL}/mcp",
      "headers": { "Authorization": "Bearer ${env:FOLIO_TOKEN}" }
    }
  }
}

[mcp_servers.folio]
url = "https://your-folio.example.com/mcp"
bearer_token_env_var = "FOLIO_TOKEN"

Claude and ChatGPT, with OAuth

In claude.ai or Claude Desktop open Customize, then Connectors, add a custom connector and paste the /mcp address. Folio shows its sign-in page and then a consent screen where you choose read, or read and write. ChatGPT takes Folio as a custom MCP server with OAuth in the same way, and Folio has the two tools, search and fetch, that ChatGPT deep research looks for. The access tokens last an hour and a refresh token renews them; you disconnect an app under API tokens, Connected apps.

Which one to pick

• Your knowledge is in Confluence Cloud or Notion and people are happy there. Use the official server. It works without a migration, and an MCP server alone is not a reason to move.
• You want Git as the source of truth, or the wiki on your own server. A Git-native wiki with its own MCP endpoint means the agent edits files, every change is a commit, and nothing leaves your host unless you send it.
• You need SSO, SCIM or user groups, or Marketplace apps. Folio has none of them yet. See Folio vs. Notion and Folio vs. Confluence for the full list of gaps.

Two smaller limits. Folio is not in Anthropic's connector directory, so you add it by address. And a write is in Git history after a short delay, not in the same second, so an agent that asks for page history right after an edit can find it empty.

Try it

The public demo runs the same endpoint, with a shared login and data that resets every 24 hours, so do not connect an agent that has a shell or your own keys; the Claude page has a prefilled connector form for it. To run your own, follow the install guide. If you are leaving a Notion or Confluence workspace, Notion alternatives and Confluence alternatives list the options with sources.

All posts