Granola MCP: Connect Claude, ChatGPT and Cursor to Your Notes

Granola's MCP server lets Claude, ChatGPT, Cursor and other AI tools search and read your meeting notes, and it works on every Granola plan, the free one included. You connect by picking Granola from a client's connector list or adding one URL, mcp.granola.ai/mcp, then signing in through your browser. No API key changes hands. The plan decides the reach. Basic exposes your personal notes from the last 30 days, without transcripts. Business, at $14 per user per month billed monthly, opens full history, transcripts and public workspace notes.^[1] On Enterprise, MCP stays off until an admin turns it on. Our verdict: start on Basic, and pay only when your AI tool needs transcripts or meetings older than a month.

What each Granola plan exposes over MCP

The 30-day line on Basic is a window, not a deletion. Older notes stay stored and reappear in the app, in Chat and over MCP once the workspace upgrades, so you can trial the connector on recent meetings before paying anything. The bigger gap is transcripts. On Basic your AI client sees the enhanced notes but cannot call the transcript or folder tools, which rules out questions that need exact wording. Business is billed monthly only, and every active member of a paid workspace is a seat; our Granola pricing guide covers both rules in detail.

For one person, full MCP access costs $168 a year by our arithmetic, billed monthly. A team of five pays $70 a month, or $840 a year. Those figures buy more than the connector, since history, integrations and the API come with the same seat. If your AI questions only ever touch the last few weeks of your own meetings, though, there is nothing to buy. Basic covers that case at $0 per user per month, with no meeting cap.

Next to other note takers' MCP servers

An MCP server is common in this category now. Otter offers one on every plan, though its public API and webhooks are Enterprise only. Fireflies runs its server at api.fireflies.ai/mcp with OAuth, and Claude Desktop can also reach it with a Fireflies API key; its three tools fetch the user, one transcript or a filtered list of transcripts. Fireflies also says its zero data retention does not cover data passed to AI tools through MCP. Krisp includes MCP on its Core plan, priced at $16 per user a month billed monthly or $8 per user a month billed annually. Granola's server accepts OAuth only, with no API-key fallback.

Before you connect: three checks

Most failed first connections trace back to something outside the AI client. Check these three points before you open Claude or ChatGPT, because each one can block the sign-in or leave the connector looking empty even when the setup itself went fine. Once they are clear, the rest is quick: Granola says setup takes under five minutes.

• Your workspace. Each connection acts for one user in one workspace at a time. If you belong to several Granola workspaces, the client reads only the active one, and a workspace-switching tool changes it.
• The Enterprise switch. On Enterprise, MCP is off by default and an admin decides whether to turn it on.^[2] The admin enables it in Settings, under Security, and chooses personal notes, public notes, both or neither. With neither chosen, nobody can connect.
• The client's own admin. In Claude Team or Enterprise, an org owner adds the Granola connector under Organization settings first. For Microsoft 365 Copilot, a Microsoft 365 admin must approve Granola before anyone connects, whatever the Granola plan.

Connecting Claude and ChatGPT

In Claude, open Customize, then Connectors, search for Granola and authenticate in the browser window that opens. Granola points Claude users to the web connector page at claude.ai/customize/connectors rather than the Claude Desktop extensions menu. The hosted server then works in Claude on the web, on desktop and on mobile, with no local config file to edit. Claude Team and Enterprise customers get two extras. Claude auto-connect for Granola is in beta on those plans. Enterprise-Managed Authorization goes further, but needs a Granola Enterprise workspace with SSO and at least 50 seats; Granola has tested it with Claude and Okta Cross App Access.

In ChatGPT, open Settings, search for Granola, connect and sign in. Each person authenticates separately with OAuth, so ChatGPT sees only the notes that person can already open in Granola. If you use Granola's shared-screen snapshots, ChatGPT can also see the slides, charts and designs captured in your meetings.^[3] Snapshots run on the Mac only (macOS 14 or later) and stay off on Enterprise until an admin enables them. Revoking is all or nothing. Disconnecting ChatGPT removes its access to the whole Granola workspace; there is no per-note switch.

Cursor, Claude Code, Codex and Copilot

Developer tools take the address by hand. In Cursor, Claude Code or any client that supports Streamable HTTP with OAuth 2.0, add the server address mcp.granola.ai/mcp (over HTTPS) as a remote MCP server.^[4] The client registers itself through Dynamic Client Registration, so you need no client ID or secret, and a browser sign-in finishes the job. Granola's Claude Code page says the agent can then list, search and read meetings and browse folders. For Codex, Granola gives two terminal commands: codex mcp add granola, then codex mcp login granola. There is no service-account method. Every connection belongs to one person.

Microsoft 365 Copilot follows an admin-first route. Granola appears in the Connectors Gallery as a partner-submitted federated connector, and once a Microsoft 365 admin approves it, Granola says each user's setup takes under five minutes. Copilot Studio can also add Granola as an MCP tool, using OAuth 2.0 and dynamic discovery. Granola says the connector neither copies nor stores data in Microsoft 365. Beyond these clients, Granola's integrations page lists Gemini, Grok, Figma, Lovable, Manus, Replit and v0, and its MCP page says any app that can use an MCP server can connect with the same address.

What your AI client can do once connected

The connector exposes a short list of tools. Your client can query meetings in plain language, list and fetch meetings, pull attachment screenshots, switch workspace and sign out; getting a transcript and listing folders need a paid plan. Granola's audit-event documentation also names tools for people and company search, meeting briefs, folder management, summary generation and note import. MCP is not only for reading. The audit log records private-note updates and summary generation done through it, and Enterprise admins can switch off changes over MCP for everyone, admins included.

In practice, the connector turns meetings into context for other work. In Cursor, an engineer can pull the decisions from last week's spec review before writing code. In Claude, a manager can draft a client update from a month of call notes. In ChatGPT, a founder can ask which customers raised the same objection across several calls. The third task leans on exact wording, so it works best on Business, where transcripts are reachable.

Mind the throughput. Granola says MCP averages about 100 requests per minute across all tools, varying by plan and tool, and subject to change. A broad prompt across a year of meetings can fan out into many tool calls, so narrow questions by date, person or folder. For bulk exports or scheduled syncs into a CRM, the API on Business is the better path, with documented rate limits and webhooks; our guide to Granola's API and integrations covers that side.

Security, revoking access and admin controls

Over MCP your AI client receives notes and, on paid plans, transcripts, but never audio, because Granola keeps none after transcription. On Granola's side, data sits on AWS in the United States, with no regional residency option. Once results reach Claude or ChatGPT, that client's own data settings govern them, so check those before you connect customer meetings. Deleting transcript segments in Granola removes them from MCP as well. That is the cleanest way to keep a sensitive passage away from an AI tool without deleting the whole meeting.

You revoke a connection inside the AI client that added it. Regenerating an MCP token stops the previous token at once. One gap matters for offboarding: removing someone from a Claude organization stops auto-connect, but a token already issued lasts until it expires. Remove the person's Granola account too. Disconnecting leaves every note in the Granola workspace untouched, and reconnecting simply means signing in again.

Enterprise adds the oversight layer. The Audit API logs MCP tokens being enabled, disabled or regenerated, plus every tool call with its outcome and the client's name and version. Admins can also turn off transcripts over MCP for the whole workspace. For a security team that wants AI access on but traceable, that is the deciding difference. MCP itself costs nothing to try; the controls around it start at Enterprise, from $35 per user per month.

The bottom line

Granola's MCP server is a low-friction way to put meeting notes inside Claude, ChatGPT, Cursor and other AI tools: a connector listing or one address, a browser sign-in, no keys. Basic is enough to test it on a month of personal notes. Move to Business when your AI client needs transcripts, older meetings or public workspace notes, and look at Enterprise only for admin scopes, the audit trail or Okta-based managed authorization. Enterprise teams should expect it switched off until an admin acts.