Guide
Connect an AI tool
Floating Keys runs a remote MCP server so an AI tool you already use can read your Projects, send messages and suggest work. You approve each tool, choose its Projects and permissions, and can revoke it at any time.
Current status#
- Public discoveryLive
OAuth discovery for the MCP server is published on floatingkeys.com.
- Connecting your accountPartial
Invited accounts only. The full sign-in, approve and tool-call journey has not been verified end to end.
- Goals and map toolsPlanned
Listed as permissions, but no tool uses them yet.
- Coding toolsPlanned
Designed, but off. Not offered during sign-in.
Two different uses of MCP#
- Floating Keys as an MCP server
- An outside AI tool connects to Floating Keys and reads or writes your Projects. This page covers that.
- MCP connectors inside Floating Keys
- Floating Keys connects out to another service’s MCP server so chat can use its tools. See Keys and connectors.
Before you connect#
- A Floating Keys account that is invited to the beta, with a verified email.
- An AI tool that supports remote MCP servers over HTTP with OAuth sign-in. Claude Code, Claude, ChatGPT and Cursor have this option; which plans include it is up to each provider.
- At least one Project, if you want the tool to see anything.
Connect a tool#
- In Floating Keys, open Settings → Integrations. Under Connected AI tools, copy the address in Connect a tool.
- Add it to your tool as a remote MCP server or custom connector. In Claude Code:
claude mcp add --transport http floating-keys <address>, then run/mcpand choose to sign in. In Cursor: add the address as a serverurlinmcp.json. - Your tool opens Floating Keys. Sign in if asked. Approving needs a sign-in from the last five minutes.
- Review the tool’s name and what it can do. Choose All projects or pick Projects, then choose Allow or Deny.
- You return to your tool, which can now call Floating Keys.
What a tool can do#
| Permission | What it allows | Status |
|---|---|---|
| See your projects | List the Projects in the grant and read their records and links (list_projects, get_project). | Partial |
| Read chats | Read the last 100 text messages of a Project chat, and see recent chat titles (read_chat). | Partial |
| Read Library | Search and read a Project’s text files (search_library). Account-wide Library only with All projects. | Partial |
| Send messages | Save a message to a Project chat and get a text reply from a free model, with no tools (send_message). | Partial |
| Suggest work | Save a Work, Code or Goal suggestion for you to review (propose_work). It never starts anything. | Partial |
| See goals | No tool uses this yet. | Planned |
| See your map | No tool uses this yet. | Planned |
A Project viewer can read but cannot send messages or suggest work. Access is checked again before each result is returned, so removing someone from a Project also removes their tool’s access to it.
Limits and uncertain results#
- A connection lasts 30 days, then the tool must ask again. A tool can narrow its permissions but never widen them.
- Each request ends after 55 seconds. Tools are limited to 120 requests a minute per account.
send_messageandpropose_worktake a request ID. Reusing it returns the same result instead of repeating the work.- If a reply’s outcome is unknown, your message may already be saved. Check the chat in Floating Keys before sending it again.
Review and revoke#
Settings → Integrations → Connected AI tools lists each tool with its Projects, number of permissions and last use. Revoke ends its access; responses in progress are closed, but text already sent cannot be taken back.
Suggested work opens on a review page in Floating Keys. Open in Work or Open in Code puts it in the composer as a draft; nothing runs until you send and approve it.
What Floating Keys keeps#
- Only fingerprints of a tool’s access codes are stored, never the codes themselves.
- An activity record keeps the action, tool, Project, time and outcome for 90 days — not your prompts, messages or replies.
- Once a tool reads something, the tool’s own provider handles it under its own terms.
Status reflects November 2026. Planned items are design targets, not commitments to a date.