Floor plan tools for AI agents
FloorGen has an MCP server with 151 tools, so Claude, ChatGPT, Cursor and other MCP clients can read and edit a real floor plan: walls, rooms, doors, furniture and code checks, not a picture of a plan. In the studio, an AI chat does the same once you sign in.
- MCP tools
- 151
- Clients
- Claude, ChatGPT, Cursor
- Auth
- OAuth or API key
- In-app chat
- With sign-in
A language model on its own can describe a house or generate an image of a plan, but the image has no scale and nothing to edit. Connected to FloorGen, the same model works on a drawing: it calls tools that create walls of a set length, place a door in a wall, move a wall and keep its neighbours joined, and read back the room areas. You open the result in the studio and see exactly what it did.
What an agent can do with a plan
| Ask | What happens |
|---|---|
| "Start a project from the US ranch template and list the rooms." | The agent creates the project from the template and reads back each room and its area. |
| "Make the primary bedroom 2 ft wider, keep the doors where they are." | It finds the shared wall and moves it; doors and windows in the wall move with it, and the room areas update. |
| "Check this plan against the IRC." | It runs the plan checks: room sizes, ceiling heights, bedroom escape openings, stairs and guards, and reports what fails. |
| "Furnish the bedrooms and the office." | It places a fitting furniture set in each room, clear of door swings. |
| "Put this sketch behind the plan and draw over it." | With the trace toolset it places the image (PNG, JPEG, WebP or GIF), scales it from two points and a real length, then draws the walls over it. |
| "Make two variants and compare them." | It makes each variant as a copy of the design, leaving the original untouched, then compares them: both plans side by side at one scale, and what each adds, removes and changes, with area totals. |
Two ways in
The MCP server at https://floorgen.app/mcp works with any client that speaks the Model Context Protocol. Apps that only take a URL, such as claude.ai, Claude Desktop's connector screen and ChatGPT, sign you in with OAuth; tools that let you set a header, such as Claude Code and Cursor, use an API key you make under Agents & MCP in your account. The MCP docs have the setup, and there are walkthroughs for Claude and ChatGPT.
The in-app AI chat sits beside the plan in the studio and edits the open project. It needs you to sign in with Google. Both routes edit the projects in your FloorGen account, so a plan you drew before signing in joins the account when you sign in.
Guard rails
An agent that can draw walls can delete them, so the server assumes it will make mistakes:
- Access levels. Read only, read & edit, or full. A read-only credential never sees the write tools; only full access can delete.
- Scoped keys. An API key can be limited to chosen projects (up to 100) and set to expire after 30, 90 or 365 days.
- Confirmation before deletes, and a checkpoint saved before an agent's first edit.
- A change log and an Activity panel in the studio, where you can see what an agent did and revert it.
The guard rails in FloorGen's MCP server explains each one and how to combine them.
A working loop
- Connect your AI app with OAuth or a read & edit key scoped to one project.
- Start from a template rather than a blank page; the agent edits a real house faster and better than it invents one.
- Ask for one change at a time, and read the areas it reports back.
- Open the project in the studio and look: the plan, the 3D view and the check results.
- Undo or revert anything you do not want, then export the plan, the print set or a walkthrough video yourself.
What FloorGen doesn't do
- It does not make the model right. The agent can misread which wall you meant. Because the result is a real drawing, you can see the mistake and undo it, but you have to look.
- It is not an architect. An agent-drawn plan is a starting point, not a permit set, and the checks cover room sizes, ceilings, escape openings and stairs, not structure, setbacks or smoke alarms.
- No image-to-plan magic. Tracing an image is semi-automatic: the scale comes from a length you confirm, and detected walls are suggestions.
- No photoreal or AI-styled renders.
- Sign-in is required for MCP and for the in-app chat. Your AI app's own plan and limits apply on its side.
Questions
Which AI apps work with FloorGen?
Any MCP client. The setup is documented for Claude Code, Claude Desktop, claude.ai, ChatGPT and Cursor; connecting Claude, ChatGPT and Cursor has the exact snippets.
Why are only some tools listed to my agent?
The 151 tools are grouped into seven toolsets. By default the server lists the 44 core tools, which keeps the list short enough for a model to choose well; you can ask for more in the server URL.
Can the agent delete my projects?
Only with full access, and only after a confirmation step. A read & edit key cannot delete anything, and a checkpoint is saved before an agent first edits a project.
Does FloorGen use my plans to train AI?
No. The terms say we do not use your content to train AI models and do not sell it. The AI app you connect has its own terms for what it reads.