FloorGen MCP server

What the server does

An MCP client calls tools such as create_project, add_walls, add_opening and place_item to build a real plan: walls with lengths, rooms that close, doors in walls. The plan is an ordinary FloorGen project, so you can open it in the editor, change it by hand and hand it back. Edits an agent makes appear in an open editor within about a second, each as one undoable step, and are listed under Activity.

Connecting a client

Clients that take only a URL (claude.ai, Claude Desktop's connector screen, ChatGPT) sign you in with OAuth: you sign in to FloorGen and pick the access level on a consent page. Clients that let you set a header (Claude Code, Cursor, VS Code and most editors) use an API key, sent as Authorization: Bearer <key>. Keys are made on FloorGen's Home screen, under Agents & MCP at the foot of the sidebar: pick Read only, Read & edit or Full, all projects or chosen ones, and an expiry. Either way you need a free FloorGen account, and only projects synced to it are visible.

Claude (claude.ai, Desktop and mobile)

1. Open Settings → Connectors → Add custom connector.
2. Name it FloorGen and paste https://floorgen.app/mcp as the remote MCP server URL.
3. Click Connect, sign in to FloorGen and choose Read only, Read & edit, or Full.
4. In a chat, turn FloorGen on from the tools menu. A connector added on claude.ai also appears in Claude Desktop and the mobile apps.

Claude Code

export FLOORGEN_API_KEY="fg_live_PASTE_YOUR_KEY"
claude mcp add --transport http --scope user floorgen https://floorgen.app/mcp \
  --header "Authorization: Bearer $FLOORGEN_API_KEY"

Check it with claude mcp list, or /mcp inside a session.

ChatGPT

1. Turn on developer mode: Settings → Apps & Connectors → Advanced settings → Developer mode. Write tools need it.
2. Create a connector named FloorGen with the MCP server URL https://floorgen.app/mcp and OAuth authentication.
3. Sign in to FloorGen when asked, pick the access level, then choose the connector in a chat and ask it to list your FloorGen projects.

Cursor

{
  "mcpServers": {
    "floorgen": {
      "url": "https://floorgen.app/mcp",
      "headers": { "Authorization": "Bearer ${env:FLOORGEN_API_KEY}" }
    }
  }
}

In Cursor Settings → MCP, floorgen shows a green dot and its tool count. The Agents & MCP dialog has set-up text for each client (Claude Code, Claude Desktop, claude.ai, Cursor, VS Code with Copilot, Windsurf, Codex CLI, Gemini CLI, Cline, ChatGPT and any other), with a new key filled in. The longer walk-through, with Claude Desktop's config file, is connecting Claude, ChatGPT and Cursor to FloorGen.

Toolsets

A short tool list helps a model choose well, so the server lists only core (44 tools) unless the client asks for more, with ?toolsets= on the URL or an X-FloorGen-Toolsets header: https://floorgen.app/mcp?toolsets=core,stairs,history, or ?toolsets=all for all 151. A credential only ever sees the tools its access level allows: a read-only key never sees an edit tool.

All 151 tools

Grouped by what they work on. Lengths are in centimetres, or text such as 3.5m or 12'6"; every edit takes dry_run and expected_rev. "Edit" tools need Read & edit access, "Delete" tools Full access and a confirmation token.

Reading the plan

Start here: what the server allows, and the plan as rows an agent can search.

The open editor tab

Work with what the person sees in the browser right now (toolset live).

Projects and settings

Create, copy, import, export and delete projects; units, code profile and title block.

Floors, designs and scenarios

Storeys, alternative designs of a floor, and whole-house scenarios.

Walls and rooms

Walls are the geometry; rooms are found from closed loops of walls.

Doors, windows and finishes

Openings in walls, curtains and blinds, floor and wall finishes.

Furniture and fixtures

Search the catalog of about 2,500 models and place, move and restyle items.

Stairs and lifts

Toolset stairs.

Surfaces, railings, notes and dimensions

Terraces, balconies, cut-outs, railings, notes, dimension lines, the north arrow and section lines.

Tracing an image or a drawing

Put a scanned plan or photo under the plan and scale it, or import a CAD drawing (DXF) as walls, doors and windows (toolset trace).

Batch edits, copies and the trash

Many edits as one undo step, moving and copying by id, and the restorable trash.

Checks and schedules

The same checks as the editor's issue list, the area schedule and geometry clean-up.

Images, elevations and the roof

Pictures the agent can look at, and the elevations the print sets use.

Comparing designs and projects

Side-by-side plans and area tables.

History, undo and checkpoints

Every agent edit is logged as a transaction that can be inspected and undone (toolset history, except undo).

One-step layouts

Bigger jobs in one call, each a single undo step with a checkpoint before it (toolset macros).

12 prompts and 15 resources

Each prompt is a playbook: a job written out step by step against the real tools, starting with get_context and ending with validate_plan and render_2d.

Safety rails

• Access levels. Read only, Read & edit, or Full (which can delete), per key or per OAuth grant. A key can be limited to chosen projects and given an expiry.
• Preview and conflicts. Every edit takes dry_run to see the result without saving, and expected_rev so an edit made on a stale copy is refused.
• Two-step deletes. Destructive tools return a confirmation token first; the agent has to ask and call again. soft_delete moves things to a trash that restore_trash empties back.
• Undo and checkpoints. Each call is one logged change that undo reverses. A checkpoint is saved before macros and deletes, and deleted projects can be restored for 30 days.
• Rate limits. 300 calls and 60 edits a minute per key.

More on how these work together: letting AI agents edit floor plans safely.