How to connect ChatGPT, Claude, Grok, and Perplexity to your notes and tasks

By Gerald · 10 June 2026

Flow does not have a built-in AI assistant. That was a deliberate choice.

Every productivity tool that ships with AI makes the same tradeoff: you pay a higher subscription to fund a model you may not trust, may not like, and may already have access to somewhere else. The AI is baked in, opinionated, and impossible to swap out.

Flow takes the opposite position. It gives you the raw data layer, structured and accessible, and lets you bring the AI model you already use and trust.

The mechanism is MCP, which stands for Model Context Protocol. It is an open standard backed by Anthropic that lets AI tools connect to external data sources and take structured actions. Think of it as a secure bridge: your AI asks for your tasks, Flow hands them over, your AI reads them, and if you approve, it can create or update them.

Every Flow account comes with an MCP connector. You copy one URL, paste it into your AI tool of choice, and your notes and tasks become available directly inside whatever interface you already use.

This article is the setup guide. If you want the honest case for whether MCP is worth connecting at all, read why use Flow MCP when copy-paste is faster first. The short version: MCP is for context, not speed.

What your AI can actually do

When your AI tool connects to Flow over MCP, it has access to a specific set of tools. These tools are intentional, not exhaustive.

Your AI can read all your boards, tasks, notebooks, notes, and captures. It can create new tasks with titles, details, due dates, checklists, tags, and priority. It can update existing tasks, including moving them between columns. It can create notes and update existing ones, including the body. It can create captures. It can search across your notes and tasks by keyword.

What it cannot do is delete anything. No note, no task, no board, no capture. That boundary is enforced in the connector, not just the interface.

Get your MCP URL

Sign in to Flow and go to your account page. Scroll to the section labelled "MCP Connector." Click "Generate connector." You will see two things: a URL that looks like https://...convex.site/mcp/flow_... and a shorter API secret that starts with flow_.

The URL has the token embedded in the path. Paste that URL directly into tools that accept an MCP server URL without separate authentication. For tools that need credentials separately, use the base URL https://...convex.site/mcp and paste the API secret as the bearer token.

If you think the token has been exposed, click "Regenerate" to invalidate the old one and create a new URL.

Setting up each tool

Claude (claude.ai)

Open claude.ai in your browser. Go to Settings, then find the Connectors or Integrations section. Click "Add connector" or the equivalent button. Paste your full MCP URL (the long one with the token in the path). Give it a name like "Flow." Save.

You do not need to enter anything in Advanced settings for Claude. The token is already embedded in the URL, so no OAuth credentials are needed. Once saved, Claude can read and act on your Flow data in any conversation where the connector is active.

Claude Desktop

Open Claude Desktop and go to Settings. On macOS this is under Claude in the menu bar, then Settings. On Windows it is under the three-dot menu. Find the Integrations or MCP Servers section. Click Add and paste your full MCP URL. Save and restart the app if prompted.

Claude Code

Claude Code reads MCP server configuration from a JSON file. Run the following to add Flow:

claude mcp add flow https://...convex.site/mcp/flow_...

Replace the URL with your actual connector URL from the account page. After adding it, Flow tools appear in your Claude Code sessions automatically. You can verify with claude mcp list.

If you prefer to configure it manually, edit ~/.claude.json and add an entry under mcpServers:

{
  "mcpServers": {
    "flow": {
      "url": "https://...convex.site/mcp/flow_..."
    }
  }
}

ChatGPT

ChatGPT supports custom MCP servers, but it requires Developer Mode to be enabled first. This feature is available on Plus, Pro, Team, Enterprise, and Edu plans.

To enable Developer Mode: open ChatGPT, go to Settings, then scroll to the Connectors section. At the bottom of that section look for "Advanced" and toggle on "Developer Mode." Once enabled, you will see an option to add a custom connector by URL.

Paste your full MCP URL (the one with the token in the path). ChatGPT uses the URL as the connection and does not require a separate API key or client secret when the token is embedded in the path. Name the connector "Flow" and save.

Standard consumer ChatGPT does not support custom MCP servers without Developer Mode. If you only have a free plan, this option is not available yet.

Perplexity

Perplexity handles MCP authentication differently from Claude and ChatGPT. It does not read the token from the URL path. Instead, it sends credentials as a separate header.

To set this up: open Perplexity and go to Settings, then Connectors or Integrations. Click Add and choose the "Advanced" option. For the URL, use the base URL without the token: https://...convex.site/mcp. For the authentication type, choose "API Key." Paste your API secret (the flow_... value, not the full URL) into the API key field.

When Perplexity connects, it sends that value as a bearer token in the Authorization header, which Flow's connector accepts.

Grok

Grok on x.ai supports MCP connectors. Go to Settings inside the Grok interface and find the Connectors section. Paste your full MCP URL. The token embedded in the URL path handles authentication, so no additional credentials are needed. Save, and Grok can immediately use your boards and notes.

Cursor

Open Cursor and go to Settings, then Features, then MCP Servers. Click the button to add a new server. Enter a name like "Flow" and paste your full MCP URL. Cursor uses MCP to give its AI coding assistant access to context outside the codebase, so you can ask questions like "what tasks are in my Executing column" or "create a note with these API details" without leaving the editor.

Windsurf and Zed

Both Windsurf and Zed support MCP servers via their settings or configuration files. In Windsurf, look under Settings for the MCP or Tools section and add your URL. In Zed, add an entry to your settings.json under context_servers:

{
  "context_servers": {
    "flow": {
      "command": {
        "path": "npx",
        "args": ["-y", "mcp-remote", "https://...convex.site/mcp/flow_..."]
      }
    }
  }
}

Any other tool that supports MCP over HTTP (streamable HTTP transport) works the same way. Paste the full URL. If it requires credentials separately, use the base URL and the API secret as a bearer token.

What this replaces

Before the MCP connector, the only way to get your Flow data into an AI conversation was copy-pasting. You copied a note, pasted it into Claude, and asked your question. Then you went back to Flow, manually typed whatever the AI suggested, and kept track of the whole thing yourself.

The connector removes that overhead for reading and updating work that already lives in Flow. Your AI can ask for the three most important tasks in your Executing column, read the details and checklists, and help you think through them without any manual transfer. When you tell it to move a task to Done, it does that too.

Copy-paste is still faster for dropping one finished paragraph into a note. I explain when each approach makes sense in why use Flow MCP when copy-paste is faster.

The data stays in Flow. The AI stays wherever you use it. The connector is the link between them.

A note on security

Your MCP token is equivalent to full read and write access to your notes and tasks. Treat it like a password. Do not paste it into a shared document, a public repository, or a tool you do not trust.

If you suspect the token has been compromised, regenerate it from your account page immediately. The old token becomes invalid the moment you do.

Flow's connector does not expose delete tools by design. This is a hard constraint, not a setting. Even if your AI tries to delete something, the connector will return an error. Your data cannot be removed through the MCP interface.

Related reading

• Why use Flow MCP when copy-paste is faster?
• Flow vs Notion: a debate about power, friction, and AI
• Why you need a second brain

Frequently asked questions

Why use MCP instead of copy-paste? Copy-paste is faster for one-off transfers. MCP is for giving your AI the full context of your notes, tasks, and projects already in Flow. I wrote the full argument in why use Flow MCP when copy-paste is faster.

Does the MCP connector work with any AI tool? It works with any tool that supports MCP over HTTP, also called streamable HTTP transport. That includes Claude, ChatGPT (with Developer Mode), Grok, Perplexity, Cursor, Windsurf, Zed, and any other tool that implements the open standard.

Does the connector send my data to third parties? The connector responds to requests from the AI tool you configure. When your AI asks for your tasks, Flow's connector retrieves that data and sends it to your AI. Flow does not send data anywhere proactively. What happens to that data inside the AI tool depends on the privacy policy of that tool.

Can I limit what the AI can read? You can mark individual notes as confidential. Confidential notes are invisible to MCP: the AI cannot list, search, read, or update them. Everything else in your account remains available. There is no per-notebook or per-board scoping yet. See why use Flow MCP for how confidential notes fit into the context story.

Can I have multiple AI tools connected at once? Yes. The connector URL works from any AI tool. You can have Claude, Cursor, and ChatGPT all pointing at the same URL simultaneously.

What happens if I regenerate the token? The old URL stops working immediately. Any AI tool configured with the old URL will fail to connect. Update each tool with your new URL after regenerating.