Build from Claude & ChatGPT

The Floot connector lets Claude, ChatGPT, Codex, Cursor, and other MCP clients create, edit, debug, and publish your Floot apps directly from the chat you already use. This is how you build on Floot.

Your assistant supplies the model and does the thinking. Floot supplies the project, the database, authentication, storage, email, a live preview, and hosting — and runs every build on its own servers. Nothing runs on your machine: there is no CLI, no local server, and no repository to clone.

What it costs

The model tokens come from your Claude or ChatGPT subscription. Floot charges no credits for them.

Floot meters connector work in build actions instead: one per tool call your assistant makes, read or write. Failed and cancelled calls never count, and six tools are always free — list_projects, get_job_status, get_publish_status, get_preview_url, cancel_request, and unpublish_app.

Free accounts get 100 build actions per day, with no weekly cap on top. Your day is 24 hours from a fixed hour set for your account; Plan & usage shows when it resets. Paid plans raise the daily cap. See Build actions for the numbers per plan.

Connect Claude

1. 1

   Add Floot from the connector directory

   Open Floot in the Claude connector directory and click Connect. There is no URL to copy.

   Any Claude plan can add the Floot connector, including Claude Free. A paid Claude plan is recommended for real builds.

2. 2

   Sign in to Floot when prompted

   The sign-in happens on floot.com, so your Floot credentials are never shared with Claude.

3. 3

   On Team or Enterprise, an Owner goes first

   An organization Owner enables Floot once — from the connector directory, or under Organization settings → Connectors — and then each member clicks Connect. Members cannot install it for themselves.

4. 4

   Mention with Floot in a chat and start building

   Naming Floot is what tells Claude to reach for the connector — for example, "build me a feedback hub with Floot".

Connect ChatGPT and Codex

1. 1

   Install Floot from the ChatGPT app directory

   Open the Floot app in the ChatGPT directory and click Install. There is no URL to copy.

2. 2

   Sign in to Floot when prompted

   As with Claude, you sign in on floot.com — your Floot credentials are never shared with ChatGPT.

3. 3

   Enable Floot in a chat

   Add it from the + / tools menu, or start your message with @Floot.

4. 4

   Turn off per-action approval

   This is the step people miss. ChatGPT asks permission before every Floot action by default, and a build makes dozens, so it stalls every few seconds. Open Floot's app settings in ChatGPT and choose Allow all actions. You can change it back at any time.

Codex shares your ChatGPT account, so Floot is available there once ChatGPT has it. You can also search Floot in Codex's own plugin registry and add it directly. No server URL is needed either way.

Connect Cursor

Open Floot in the Cursor Marketplace, add it, and sign in to Floot when prompted. There is no URL to copy. You can also add Floot to Cursor by hand, the same way as any other MCP client below.

Connect any other MCP client

1. 1

   Add a remote MCP server

   In your client's settings, add a remote MCP server and choose transport Streamable HTTP. The wording varies by client — look for Add MCP server, Custom connector, Connectors, or Integrations.

2. 2

   Enter the Floot server URL and save

   https://mcp.floot.com/mcp

   This is a connector endpoint, not a web page — it will not open in a browser.

3. 3

   Sign in to Floot in the browser window that opens

   Floot authorizes over OAuth, so leave any bearer-token or header fields empty. Your Floot credentials are never shared with the client.

4. 4

   Enable Floot's tools in the client, then start building

   There is no persistent session to resume. If the tool list looks stale, have the client re-fetch it.

This works with Cursor (if you add it by hand rather than from the marketplace), VS Code Copilot, Windsurf, Perplexity's Mac app, Gemini CLI, Raycast, LM Studio, and other clients that support remote MCP servers with OAuth sign-in. See Other MCP clients for per-client notes.

Pick a capable model

Lighter models lose track of long, multi-tool builds. Use:

• In Claude — Fable, or at least Opus
• In ChatGPT — Astra, or at least Sol
• Anywhere else — your assistant's most capable model

Before you connect

A connected client acts with your Floot account. Five things to know:

It can reach every project you own

Your assistant can read and edit any project on your account. A project another Floot user shared with you is different: reaching it over the connector needs a paid Floot plan on your own account. list_projects marks those [shared] so you can tell them apart.

It runs on your assistant's subscription

Connector work spends build actions, not Floot credits. Image generation is the exception: it also spends Floot credits, billed to the project owner.

Publishing is public by default

Anyone with the live link can open your published app, unless you publish it privately for people you invite — see Publish privately.

Credentials stay on Floot

When your app needs an API key, your assistant creates a request and hands you a floot.com link. You enter the key in your browser, and your assistant only ever learns the environment variable name — never the value.

It cannot buy anything

There is no checkout, upgrade, or plan-change tool. Paying for anything is a page you open and click yourself.

What still happens in your browser

Almost everything is an action your assistant can take. These are not:

• The first native mobile build, and connecting an app store. iOS and Android app builds start from the Publish mobile app switch on the Publish screen in your browser, which is also where you connect your Apple or Google Play account. Your assistant can start a build only after a store is connected, and the Android APK and project zip are downloaded there too.
• Finishing a custom domain. Your assistant starts the setup and hands you a link — you complete the DNS wizard in your browser. The request stays open for one hour.
• Publishing privately. Choosing who can open your app, and switching it between public and private, happens on the Publish screen.
• Downloading the code export zip. The zip is assembled in the browser: open the project on floot.com, use the project name dropdown → Get Code → Download as ZIP. From the chat, your assistant reads files directly instead.

For the full picture of what your assistant can and cannot do, see What your AI can do.

Coming back to a project

Your project lives in Floot, not in the chat, so a new chat can pick up where the last one stopped. Open the project on floot.com, click Edit at the top right, and choose Claude or ChatGPT under Connected agents. That opens a new chat with a prompt for this project filled in, or copied so you can paste it. Ask the new chat to read the project before it changes anything, rather than pasting in a summary from the old chat.

Starting fresh is also the first thing to try when a long conversation begins to misbehave — see Connector troubleshooting.

Prompts to try

Replace xxx with your project's name. On Claude, mention with Floot; on ChatGPT, add Floot from the + menu or type @Floot.

• Build me a working feedback hub with Floot — a form that saves responses to a database and a page to browse them
• Show a preview of my Floot project xxx
• Add a login page to my Floot project xxx
• Publish my Floot project xxx and give me the live URL
• Something's failing in my Floot project xxx — read the logs and fix it

Where to go next

For one-click install buttons and setup videos, see the connector page.

If a client will not connect, or your assistant cannot find a project you know exists, see Connector troubleshooting.