Connector troubleshooting

Fixes for the things that go wrong connecting Floot to Claude, ChatGPT, or another MCP client — and the behaviour that only looks broken.

Most connector problems are one of a short list of known things, and several of them are working as intended. Read down this page before assuming something is broken.

Claude will not let me add the connector

Adding connectors requires a paid Claude plan. On Claude Free there is no way to install Floot — the option is not there to find. Any paid Claude plan works.

On Claude Team or Enterprise, nothing happens

An organization Owner has to enable Floot once for the whole organization, from the Claude connector directory or under Organization settings → Connectors. Each member then clicks Connect for themselves. A member cannot self-install before the Owner has enabled it.

ChatGPT asks permission before every action

That is ChatGPT's default for connectors. A Floot build makes many tool calls, so the approval prompt reappears every few seconds and the build reads as stalled. Open Floot's app settings in ChatGPT and choose Allow all actions. You can change it back at any time.

ChatGPT only searches — it will not edit anything

Standard ChatGPT can use only two tools from any connector, and they have to be named search and fetch. Floot ships both so the connector installs and works there at all, but those two only search and read your projects. It is not a Floot restriction — Floot registers the same full tool set for every client and holds nothing back from ChatGPT. Which of them ChatGPT puts in front of the model is decided on ChatGPT's side. If you need writes and edits and ChatGPT will not make them, use a client that calls the full set — see Connect any MCP client.

My assistant does not seem to know about Floot

Enabling a connector is usually per chat. In Claude, mention with Floot in your message. In ChatGPT, add Floot from the + or tools menu, or start your message with @Floot. In other clients, switch Floot's tools on for the chat. If it still does not appear, check the connector is enabled in the client's own settings and start a new chat.

The server URL will not open in my browser

It is not meant to. https://mcp.floot.com/mcp answers MCP requests, not browser requests. It goes in your client's MCP server settings — see Connect any MCP client.

My client is asking for an API key or bearer token

Leave those fields empty. Floot authenticates with OAuth: your client opens a Floot page in your browser and you approve the connection there. There is nothing to paste, and a value typed into a token field is not a credential Floot accepts. Floot also requires PKCE with S256, so a client that offers only plain or no PKCE cannot connect at all.

Sign-in ends on a Floot page instead of back in my chat

That is the normal end of the flow. Consent runs on floot.com so your Floot credentials stay on Floot, and the page you are left looking at is Floot's, telling you to go back and enable Floot in your chat. Nothing has gone wrong — switch back to your assistant and start building.

The connection drops, or the tools disappear

There is no session to lose. Floot's MCP server holds no state between calls and issues no session ids — every request carries its own token and stands alone. When a client reports a lost connection, what it actually needs is to list Floot's tools again, which reconnecting or reloading the client does. Nothing on Floot's side is holding stale state.

My assistant says it cannot publish

It can. publish_app, republish_app, and unpublish_app are real tools, and publishing through them runs the same code as the Publish button in the Floot web app. If your assistant insists publishing has to happen in the browser, ask it directly to call publish_app. The two things it genuinely cannot do are turning on native mobile app builds and finishing a custom-domain setup — see What your AI can do.

A publish never finishes

A production build is capped at 15 minutes of wall clock. Past that the status check returns a definite failure rather than reporting the build as still running, so your assistant stops polling and tells you instead of waiting forever. Ask it to publish again.

Actions are being refused

You have hit a build-action limit. The refusal names which limit and when it resets, and status polling, list_projects, get_preview_url, cancel_request, and unpublish_app keep working meanwhile. See Build actions.

A project someone shared with me will not open

Opening a project owned by another Floot user over the connector needs a paid Floot plan on your own account, for reading as well as writing — unless the project is public, which stays readable by anyone. Projects you own yourself are never affected.

Use your client's strongest model

Lighter models lose track of long, multi-tool builds, and the symptoms read like a broken connector — steps skipped, files half-written, the same call repeated. In Claude use Opus or a more capable model; in ChatGPT use GPT-5.5 Sol or a more capable model; elsewhere pick your assistant's most capable model.