Connect your agent

No terminal? Connect from claude.ai, Claude Desktop or ChatGPT

Start here if you use Claude or ChatGPT in the browser or the desktop app. You sign in with your Flowproof account, so there is no API key to make or keep safe.

1. Open the connector settings. In claude.ai or Claude Desktop, open Settings, then Connectors. In ChatGPT, turn on Developer mode first (Settings, then Security and login), then open Plugins.
2. Choose Add custom connector. In ChatGPT, press the plus button.
3. Name it Flowproof and paste this address: https://mcp.flowproof.dev/mcp
4. Add it, then connect it. Sign in with your Flowproof account when asked.
5. Press Allow.

Then start a chat and ask it to test. The app now uses Flowproof as you. You can see it, and revoke it, under Connected apps on the Setup page.

Menu names change often, and some plans hide custom connectors: ChatGPT offers Developer mode on Plus, Pro, Business, Enterprise and Education, on the web, and on a team plan an owner may have to allow it first. Claude Code, Cursor and Codex don’t sign in like this. They use an API key, below.

What you see in claude.ai: a Flowproof card

When a run ends, or you ask for a report, apps that support it show a Flowproof card right in the chat. Apps that don’t, such as Claude Code in a terminal, get the same answer as plain text.

• The grade. A stamp and a score, then the top fixes.
• Get to A+. The steps from your grade to the top one. Each has a Do this button that asks your AI to do it.
• Proven. What the run showed works, with “earlier run” on anything carried forward from one.
• Open in dashboard for the full report.
• Share. A public link to the score card: the grade, the fixes and what was tested. Never the test addresses, email contents or project name. Stop sharing from the card, or from the run in your dashboard, and the link stops working at once. You can also just ask your AI to share the report.

After cleanup, a second card, the receipt, says what was deleted and from where: “Deleted from DrawRight: 1 user, 1 profile.”

Chat apps often can’t drive a browser. When a step means using your app, your AI tells you what to click and you click. Flowproof doesn’t count your typing time against your app’s email speed.

Claude Code

Terminal

claude mcp add --transport http flowproof https://mcp.flowproof.dev/mcp --header "Authorization: Bearer fp_YOUR_KEY"

Cursor

Editor

{
  "mcpServers": {
    "flowproof": {
      "url": "https://mcp.flowproof.dev/mcp",
      "headers": {
        "Authorization": "Bearer fp_YOUR_KEY"
      }
    }
  }
}

Windsurf

Now called Devin Desktop

{
  "mcpServers": {
    "flowproof": {
      "serverUrl": "https://mcp.flowproof.dev/mcp",
      "headers": {
        "Authorization": "Bearer fp_YOUR_KEY"
      }
    }
  }
}

VS Code

With GitHub Copilot

{
  "servers": {
    "flowproof": {
      "type": "http",
      "url": "https://mcp.flowproof.dev/mcp",
      "headers": {
        "Authorization": "Bearer fp_YOUR_KEY"
      }
    }
  }
}

Codex

OpenAI’s Codex CLI, in the terminal

[mcp_servers.flowproof]
url = "https://mcp.flowproof.dev/mcp"
http_headers = { "Authorization" = "Bearer fp_YOUR_KEY" }

Claude Desktop

Chat app

{
  "mcpServers": {
    "flowproof": {
      "command": "npx",
      "args": [
        "-y",
        "mcp-remote",
        "https://mcp.flowproof.dev/mcp",
        "--header",
        "Authorization:${AUTH_HEADER}"
      ],
      "env": {
        "AUTH_HEADER": "Bearer fp_YOUR_KEY"
      }
    }
  }
}

Replit

Replit Agent, in the browser

https://mcp.flowproof.dev/mcp

Bearer fp_YOUR_KEY

Other tools

Any MCP client

{
  "mcpServers": {
    "flowproof": {
      "type": "http",
      "url": "https://mcp.flowproof.dev/mcp",
      "headers": {
        "Authorization": "Bearer fp_YOUR_KEY"
      }
    }
  }
}

https://mcp.flowproof.dev/mcp

Bearer fp_YOUR_KEY

Building in Lovable or Bolt?

Flowproof needs an agent that can run your app and sign up to it. Open your project’s code in one of the tools above and connect Flowproof there.

Then ask it to test

Once your tool shows Flowproof as connected, pick where your app is and paste a prompt into its chat.

• My app is built. Your agent opens a test run, signs up with throwaway addresses, checks every email, and deletes the test users at the end.
• I’m building my app. Your agent plans the emails your app needs (who gets each, why it matters, what it must have), saves them to the email map as planned, then builds and tests each one with you as you go.

Runs appear on the Runs page of your dashboard.

Keep every email to the bar

Paste these rules into the file your AI tool reads at the start of every session. From then on, whenever your agent adds or changes an email, it adds it to the email map, sends it through your real provider, includes a plain-text part, names your product in the subject and first line, adds the “didn’t ask for this?” line on security email, and tests it with Flowproof before calling the task done. Your agent can also fetch them itself: ask it to add Flowproof’s email rules.

<!-- flowproof:email-rules -->
## Emails: built and tested with Flowproof

This app’s emails are planned, built and tested with Flowproof (the `flowproof` MCP server). Whenever you add or change code that sends email:

- When you add or change an email, add it to the Flowproof email map with `save_email_map`, status `built`. If it is in the map as `planned`, save the same key as `built`.
- Send it through the app’s real email provider. Never leave a `console.log` stub, a fake transport or a TODO in a code path production runs.
- Include a plain-text part as well as the HTML.
- Name the product in the subject and in the first line.
- On security email (verification, sign-in links and codes, password resets), add a line for people who didn’t ask for it: “Didn’t ask for this? You can ignore this email.” On notices that a password, email address or sign-in changed, say what to do if it wasn’t them.
- Give each email what its kind must have: `plan_emails` lists it (a reset says how long the link lasts and works once; a failed payment shows the amount and links to update the card; an invite says who invited them to what).
- Before you call the task done, test the email with Flowproof: `whats_next` first (it says where the project stands and what to do), then `start_session`, `new_address` for each recipient, `mark_step`, trigger the email, `wait_for_email` with its `map_key`, use its link or code and `record_outcome`, then `end_session`. Tell the user the grade and what to fix.
- Re-test only the emails whose map entry changed in this task. The others keep their results.
- Never test with a real inbox: use only addresses from `new_address`, and open the session before creating any user.
- Build a `suggested` email only after the developer accepts it (`accept_suggestions` makes it `planned`). Every email that is not transactional has an unsubscribe link and a link to email preferences; marketing goes only to people who opted in.
<!-- /flowproof:email-rules -->

The Get started checklist

Your dashboard’s home page has a five-step checklist: create a key, connect your agent, run your first test, connect cleanup, and map your emails. Each step ticks itself from what has actually happened (a key your agent has used, a finished run, an armed cleanup connection, a saved email map), never from a button you click.

It hides itself once every step is done. To see it again, open your dashboard with the checklist showing.

If it doesn’t work

• Your tool doesn’t list flowproof. Check the file is in the place the steps say and is still valid JSON (a missing comma between two servers is the usual culprit), then restart the tool.
• It’s listed but says it can’t connect, or the key isn’t recognised. Check you replaced fp_YOUR_KEY with the whole key, including fp_, kept the word Bearer and the space after it, and that the key isn’t revoked on the Setup page.
• Claude Desktop: “npx” not found. Install Node.js from nodejs.org, then quit and reopen Claude Desktop.
• You lost the key. Keys are shown once and can’t be shown again. Revoke it and make a new one; it takes a click.
• Connected, but the agent can’t test your app. It needs a running copy it can use: on your computer, a staging site or production. See the quickstart for more.