Connect any MCP client

Floot runs a remote MCP server. Any AI client that can add a remote MCP server over Streamable HTTP can create, edit, run, and publish Floot apps. You add one URL and sign in to Floot in your browser.

Claude and ChatGPT need none of this — both install Floot in one click from their own directories. See Build from Claude & ChatGPT. This page is for every other client.

Server details

Add Floot to your client

1. 1

   Open your client's MCP settings

   Look for MCP servers, Connectors, Custom connector, or Integrations. The wording differs from client to client.

2. 2

   Add a remote server and choose Streamable HTTP

   If your client asks for a transport, pick Streamable HTTP. Floot answers in JSON — there is no SSE stream to select.

3. 3

   Enter the server URL and save

   https://mcp.floot.com/mcp

   Leave any bearer-token, API-key, or header fields empty. Floot signs you in with OAuth, so there is nothing to paste — and a value typed into those fields is not a credential Floot accepts.

4. 4

   Sign in to Floot to authorize the connection

   Your client opens a Floot page in your browser. Approve the connection there. Because sign-in happens on floot.com, your Floot credentials are never shared with the client.

5. 5

   Enable Floot's tools and start building

   Some clients switch a new server's tools on automatically; others need you to enable them per chat. Mention Floot in your first message.

How the sign-in works

Floot's MCP server is its own OAuth authorization server, so most clients need no configuration beyond the URL.

• OAuth 2.1 authorization-code flow with PKCE. `S256` is the only challenge method Floot accepts — a client that offers only plain, or no PKCE at all, cannot connect.
• Dynamic client registration is supported, so your client registers itself. There is no client id or secret for you to create.
• The consent screen is https://floot.com/mcp-authorize, which is where your Floot session already lives.
• Access tokens are sent as a normal Authorization: Bearer header by the client, after the OAuth flow — not by you.

Clients this works with

Floot's own connect instructions name these as clients that support remote MCP servers with OAuth sign-in: Cursor, Perplexity (Mac app), the Gemini CLI, VS Code Copilot, Windsurf, Raycast, and LM Studio. That is a list of examples, not a whitelist — anything that speaks remote MCP over Streamable HTTP connects the same way. Codex is the exception that does not belong here: it shares your ChatGPT account, so Floot is already there once ChatGPT has it, and it has its own plugin registry — see Build from Claude & ChatGPT.

Floot registers the same tools for every client. Nothing is held back from a client because of what it is called. See What your AI can do.

After you connect

Use your client's most capable model

Lighter models lose track of long, multi-tool builds. Floot's guidance for any client other than Claude and ChatGPT is simply to use your assistant's most capable model.

There is no session to keep alive

Every request stands on its own — Floot's MCP server holds no state between calls and issues no session ids. If your client says it lost the connection, all it needs is to list Floot's tools again.

Building is metered in build actions

One per tool call your assistant makes, on every client. See Build actions.

Discovery

https://floot.com/.well-known/mcp.json describes the server — endpoint, transport, and auth type — for clients and registries that look it up. Floot is also listed in the Model Context Protocol registry as com.floot/floot.

If a step above does not go the way it reads here, see Connector troubleshooting.