← BLOG
· Arc0 team

How to connect Claude to Notion and limit what it edits

Connect Claude to Notion three ways, see which pages it can reach, switch off its write tools or hold edits for approval, plus test prompts and fixes.

You connect Claude to Notion so it can find last quarter's planning doc. Then you notice it can also edit the team wiki, and nobody said which pages, or whether "edit" includes moving them to the trash.

Short answer: Notion's own connector is the quickest way to connect Claude to Notion, and it's fine if you mostly read. To block edits outright, switch Notion's write tools off in Claude. For rules that follow you to other clients and a log of every call, use a custom connector that enforces them. Arc0, which publishes this blog, is that route and the worked example below. On its free Personal plan, trashing waits for your yes once you set it to ask, and holding every edit as well is one policy line on Build, which is also free.

Three ways to connect Claude to Notion

Every route ends the same way: Claude talks to Notion over MCP (Model Context Protocol), the standard way an assistant calls outside tools. The difference is who runs the server on the other end, and whether anything stands between Claude's enthusiasm and your wiki.

1. Notion's connector from Claude's directory. Claude's directory has a Notion connector that searches, creates, and updates pages, and it points at Notion's hosted server, https://mcp.notion.com/mcp. To add it, select + in a chat, then Connectors > Manage connectors (or go to Customize > Connectors), click Notion, and choose Connect. On Team and Enterprise, an Owner first adds it under Organization settings > Connectors > Browse connectors with Add to your team. This suits you if you live in Claude on the web or Desktop and mostly want answers out of Notion.

2. Notion's URL in Claude Code. Claude Code adds the same hosted server from the terminal:

claude mcp add --transport http notion https://mcp.notion.com/mcp

Then run /mcp and finish the OAuth sign-in in your browser. OAuth is the consent screen where you approve what an app may do on your behalf. The --scope flag decides where the server lives: local (the default, this project only), project (shared through .mcp.json), or user (every project you open). This suits you if Claude Code is where your Notion questions come from, such as "what did the spec say about retries?"

3. A custom connector pointing at another MCP URL. Claude lets you add any remote MCP server under Customize > Connectors, with + Add > Add custom connector, a name, and the server's URL. Anthropic doesn't verify custom connectors, so only add servers you trust. This suits you if you want rules, a log, or the same setup in more than one client, which is where Arc0 comes in.

If Claude only ever reads your Notion, start with Notion's connector and stop there. One aside: Claude allows one custom connector on the Free plan, so make it a URL that carries more than one app.

Plans and admin rights

Two sets of rules apply, Claude's and Notion's, and either one can stop you.

RequirementClaudeNotion
Who can connectCustom connectors work on Free (one only), Pro, Max, Ultra, Team, and EnterpriseWorkspace members only; guests are refused
Who adds it for a teamOn Team, only Owners and Primary Owners add custom connectors; on Enterprise, also custom roles with Manage access to Libraries. Members then choose Connect themselvesAdmins can restrict adding MCP servers
Organization controlsTeam and Enterprise Owners set tool permissions per connector (see below)Approving which MCP clients the workspace allows is Enterprise-only

Notion also offers MCP tokens for connecting without a browser sign-in, but they need Business or Enterprise and a workspace owner.

If you're a guest in someone else's workspace, no route on this page gets you past that second column. Ask to be made a member.

What Claude can see in Notion

Which pages Claude sees is decided in Notion, by one of two access models. Knowing which one you're on explains most "why can't it find my page?" moments.

Notion MCP acts as you. When you connect Notion's hosted server, it works with your permissions and, in Notion's words, "can access everything you can access." So the HR page you forgot you could open is in scope, right next to the wiki.

Connections see only what's shared with them. Notion's other model is a connection. An internal connection, which a workspace owner creates, sees no pages until someone shares them through the page's ••• menu > Add connections. A public connection signed in with OAuth sees the pages you pick in Notion's page picker, and sharing a parent page includes its children. Here pages go missing because nobody shared them.

Diagram comparing Notion's two access models: on the left, Claude reaches Notion MCP, which acts as you and reaches every page you can open; on the right, Claude reaches a connection that sees only the pages shared with it, so unshared pages are missing

"Which pages" is Notion's job, and it's a blunt one. "Which actions" is the lever you actually want: a read changes nothing in your workspace, while a write or a trash does. Reads still expose whatever they reach, which is why that HR page is worth a second look.

Keep Claude from editing everything

You get two layers of control: the one inside Claude, and one that sits above it.

Inside Claude. You can switch a connector's tools off in the Search and tools menu and review the approval prompts Claude shows you. That's the simplest way to block edits today: switch Notion's write tools off in that menu, and Claude can't call them, however persuasive the prompt. Anthropic's advice is to use Allow always only for tools you trust. Watch Research mode in particular: during Research, Claude can call connector tools without asking again, so switch Notion's write tools off before you start one.

On Team and Enterprise, Owners can go further. Under Customize > Connectors, each connector has Tool permissions with Always allow, Needs approval, or Blocked per tool, applied across the organization, and users can't override them.

That's solid, with one catch: the controls live in Claude, so they don't follow you to ChatGPT or Cursor. And any approval prompt, Claude's or Arc0's, is a habit rather than a rule if you click past it at the end of a long day. A switched-off tool is a rule.

Above Claude. The other layer is a policy check that runs between the assistant and Notion, whichever client made the call. Arc0 is the example.

Set it up with Arc0 (one URL, graded actions, rules)

Arc0 is one MCP URL that reaches Notion and the rest of a catalog of 1,484 apps, with every action graded read, write, or destructive. The Personal plan is free, and the steps below set it up.

  1. Get your URL. Sign up through Arc0 for individuals. Your personal MCP URL sits under My apps.

  2. Connect Notion in Arc0. Add Notion from My apps. Which pages it can see is still Notion's decision, as in the section above.

  3. Add the URL to Claude. In Customize > Connectors, choose + Add > Add custom connector, name it Arc0, paste https://api.arc0.ai/mcp, review the authentication setting, and choose Add. Then sign in to Arc0, select your workspace, review the access, and approve. Adding the URL alone doesn't finish setup.

  4. Or add it to Claude Code. Run the command below, then /mcp, select Arc0, authenticate, pick your workspace, and review the access. Check it with claude mcp get arc0.

    claude mcp add --transport http --scope user arc0 https://api.arc0.ai/mcp
  5. Choose apps for the assistant. Under My agents, open Choose apps and keep Notion on for Claude, and off for any assistant that has no business in your wiki.

The MCP client docs and the Claude setup page cover the same steps.

The ten Notion actions

Arc0 exposes ten Notion actions, and nothing else:

ReadWriteDestructive
notion.searchnotion.create_pagenotion.trash_page (move a page to trash)
notion.retrieve_pagenotion.append_blocks
notion.get_page_contentnotion.update_page_properties
notion.get_page_property
notion.retrieve_data_source
notion.query_data_source

What's missing matters as much: no permanent delete, no comments, and no file actions. If you need Claude to leave comments, Notion's own connector is the better fit for that job. For a workspace you're nervous about, a short list is the point. The two edit actions can still change pages you care about, and that's what the next rule is for. See the Notion app page for each action's details.

The rule that holds the trash button

On Personal, the guardrail covers destructive actions: before notion.trash_page runs, Arc0 can ask you, allow it, or block it. Set it to ask. The safest delete button is one that waits for you.

That guardrail doesn't hold writes. Policies do, and they come with Build rather than Personal: Build is also free, includes MCP, policies, and approvals, and you start it from sign up. Policies set Allow, Ask, or Deny per scope (read, write, destructive), as an app rule or an override for one action, and the most specific rule wins. The docs' example is stripe:write~ask; the same pattern for Notion looks like this:

notion:write~ask
notion:destructive~ask

Now the database row Claude "helpfully" updated on Friday afternoon waits for you on Monday instead. An approval holds the exact request Claude made, and the agent can't approve its own. Fail closed is a setting you opt into, not the default. The policies docs and approvals docs have the full syntax.

Diagram titled The rule sits between Claude and your workspace: Claude's request goes to a policy check inside Arc0, where your rules decide, and a dashed line runs from Claude to an activity log of allowed, held, and blocked calls; reads exit the check to Notion, while writes and notion.trash_page exit to Held for you, which holds trash and Build writes once they are set to ask, and an approve arrow runs from Held for you on to Notion; a footnote names the three write actions, notion.create_page, notion.append_blocks, and notion.update_page_properties

The same URL in ChatGPT and Cursor

The rule lives in Arc0, not in Claude, so it applies, per agent, to calls from any client you connect to the URL, including ChatGPT, Cursor, Claude Code, and Codex. In ChatGPT, open Plugins, select the plus button, choose Add custom MCP server, name it Arc0, paste the URL, pick OAuth, and select Create as a plugin. The ChatGPT setup page has the details, and the Cursor setup page has Cursor's steps.

That's one more step than Notion's connector, and it's the step where the rules live.

Check what ran

Read the log after the first week, not after the incident. In Arc0, Activity has three tabs, All calls, Allowed, and Held or blocked, and you can filter by agent and app and see why each call landed where it did.

The audit log is on every plan: Personal keeps it for 7 days and Build for 30. Request and response payloads aren't stored by default, and credentials never reach the model. The audit log docs and the security page explain what's kept.

On the Notion side, Enterprise workspaces let admins manage which MCP clients are approved under Settings > Notion MCP > Manage workspace clients. If your company is on Enterprise, ask your admin there before you connect, not after.

Three test prompts

Start with a read, check Activity for the call, and only then try a write. Each prompt below names the action Activity should show, so you can tell a good run from a lucky one.

  • Find a page. "Find our onboarding checklist in Notion and give me its link." Activity should show notion.search, then notion.retrieve_page or notion.get_page_content to read it. A good result names the page you meant, not a near-namesake from an old archive.
  • Summarise a database. "Summarise the open items in the Product roadmap database, grouped by owner." Activity should show notion.query_data_source. A good result uses the database's own fields; if it invents a column, it didn't read the schema.
  • Draft a page. "Draft a page called Meeting notes under Team wiki with headings for decisions and follow-ups." Activity should show notion.create_page. On Personal the call runs and lands under Allowed; on Build, a notion:write~ask policy holds it for your approval, which is exactly what you're testing.

If the first prompt comes back empty, start with the first fix below.

Troubleshooting

Menus, plan rules, and limits in this post are as checked on November 10, 2026, and both apps do move them around. Beyond that, most breakages are a page nobody shared, a sign-in that went stale, or a workspace asking Notion for too much at once. None of them is Claude being clever.

  • Pages missing from results. The page probably isn't shared with whatever is asking. With a connection, share it through ••• > Add connections; with the OAuth page picker, run the picker again and include the page or its parent.
  • Notion sign-in broke in Claude Code. Run /mcp, select the server, and choose Clear authentication, then sign in again. To remove it for good, run claude mcp logout notion before claude mcp remove notion so the grant is revoked, not just forgotten.
  • Notion sign-in broke anywhere else. Notion's advice is to disconnect and reconnect, then check your Notion permissions.
  • Arc0 connected to the wrong workspace. Revoke the agent's authorization in Arc0 and authorize again with the right one.
  • Guest account or an enterprise policy error. Guests can't connect Notion MCP, and a policy error means an admin has restricted MCP servers, so ask them.
  • Rate limits on a large workspace. Notion allows 600 requests a minute per connection on Business and Enterprise and 180 on other plans, with a separate limit on database queries and a workspace-wide one shared by every connection. If Claude reports a rate limit, wait and try again. For builders: Notion returns HTTP 429 with a Retry-After header, and Notion MCP hands back retry_after_seconds. Arc0's clients never retry a write automatically, so a throttled create doesn't quietly become two pages.

A summary of a big database is where you'll hit the query limit first. Ask Claude to narrow by a filter before you ask it to read everything.

So which route is yours? If Claude only reads, Notion's connector is all you need, and switching its write tools off is the quickest way to block edits. If you want the same rules in ChatGPT and Cursor, every call logged, Claude limited to ten graded actions, and trashing held for you once you set it to ask, use Arc0, which is free to start.

To set it up, get a free personal MCP URL from Arc0 for individuals, connect Notion, and add the URL to Claude. Then ask it to find one page.

Get started

One plug. Every app. Any agent.

Your users connect their apps once, under your brand. Your agent, or Claude, ChatGPT and Cursor, acts through one MCP endpoint, with every action on the record.