scandrop

Minimal but real MCP server that ingests a GLB/GLTF floorplan scan (for example captured with Polycam https://poly.cam/) and exposes deterministic spatial tools over a compact derived scene graph.

Install

uv sync --python 3.12

Claude Desktop setup

Use scandrop in combination with the filesystem MCP so Claude can both run spatial tools and read your local .glb/.gltf files.

After uv sync, configure Claude Desktop (claude_desktop_config.json) with both servers:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/Users/<username>/Downloads", "/Users/<username>/Desktop"]
    },
    "scandrop": {
      "command": "uv",
      "args": ["run", "--project", "/absolute/path/to/scandrop-mcp", "python", "-m", "scandrop.main"]
    }
  }
}

Restart Claude Desktop after saving the config.

What scandrop MCP does

• Loads .glb / .gltf mesh exports (including Polycam free-plan outputs).
• Derives a compact spatial database:
  • dominant floor plane
  • floor polygon in XZ
  • obstacle footprints + OBBs + heuristic semantic labels
  • scene AABB bounds
• Persists versioned scene artifacts on local disk:
  • ./data/scenes/{scene_id}/v{version}/scene_graph.json
  • ./data/scenes/{scene_id}/v{version}/params.json
• Exposes deterministic MCP tools (check_fit, find_free_spaces) so the model never does geometry math.
• Provides optional FastAPI debug endpoints.

Quickstart

1) Ingest a scan

python scripts/ingest.py path/to/scan.glb

Example output:

{
  "scene_id": "ab12cd34ef56",
  "version": 1
}

2) Run MCP server (stdio)

python -m scandrop.main

2b) Configure MCP clients (e.g., Claude Desktop)

After uv sync, most MCP clients can launch this server via uv run from the project directory. In Claude Desktop, keep this server enabled together with filesystem.

Claude Desktop config (claude_desktop_config.json):

{
  "mcpServers": {
    "scandrop": {
      "command": "uv",
      "args": ["run", "--project", "/absolute/path/to/scandrop-mcp", "python", "-m", "scandrop.main"]
    }
  }
}

Generic MCP client config shape:

{
  "servers": {
    "scandrop": {
      "command": "uv",
      "args": ["run", "--project", "/absolute/path/to/scandrop", "python", "-m", "scandrop.main"]
    }
  }
}

2c) Load a scene (recommended user flow)

In Claude Desktop, users only need to provide a path to the 3D model:

Please onboard this 3D model: /Users/<username>/Downloads/test.glb

Scandrop handles scene creation, status, and summary internally via onboard_scene.

3) Run FastAPI debug server (optional)

python scripts/run_api.py --reload

Claude Desktop tutorial (attach + ingest a GLB)

Use this flow when you have both connectors enabled:

• filesystem connector (to read files from your machine)
• scandrop MCP server (this repo)
• A GLB/GLTF scan export (for example from Polycam)

1. Place your .glb file in a folder exposed by filesystem, for example /Users/<username>/Downloads/test.glb.
2. Send this prompt:

Please onboard this 3D model: /Users/<username>/Downloads/test.glb

1. The assistant runs onboarding in one step and returns:

• scene_id and version
• processing status (usually ready)
• scene summary (bounds, floor area, obstacle count)

1. Ask a placement question, for example:

Where do I fit a bedside table by the bed?

For the full web app tutorial and UI screenshot, see web/README.md.

MCP tools

create_scene_from_gltf

Input:

{ "path": "/absolute/path/to/scan.glb" }

Output:

{ "scene_id": "ab12cd34ef56", "version": 1 }

onboard_scene

Input:

{ "path": "/absolute/path/to/scan.glb" }

Output:

{
  "scene_id": "ab12cd34ef56",
  "version": 1,
  "status": { "status": "ready", "message": null },
  "summary": {
    "bounds_aabb": { "min": [0.0, -0.1, -2.0], "max": [5.2, 2.8, 4.4] },
    "floor_area_m2": 17.42,
    "obstacle_count": 6
  }
}

get_processing_status

Input:

{ "scene_id": "ab12cd34ef56" }

Output:

{ "status": "ready", "message": null }

get_scene_summary

Input:

{ "scene_id": "ab12cd34ef56", "version": 1 }

Output:

{
  "bounds_aabb": { "min": [0.0, -0.1, -2.0], "max": [5.2, 2.8, 4.4] },
  "floor_area_m2": 17.42,
  "obstacle_count": 6
}

get_scene_graph

Input:

{ "scene_id": "ab12cd34ef56", "version": 1 }

Output:

{
  "units": "meters",
  "room": { "floor_plane": { "normal": [0, 1, 0], "d": -0.02 }, "floor_polygon_xz": [[0, 0], [5, 0], [5, 4], [0, 4]], "bounds_aabb": { "min": [0, -0.1, 0], "max": [5, 2.7, 4] } },
  "obstacles": [{ "id": "obs_001", "footprint_xz": [[1, 1], [1.8, 1], [1.8, 1.7], [1, 1.7]], "obb": { "center": [1.4, 0.5, 1.35], "extent": [0.8, 1.0, 0.7], "rotation": [[1, 0, 0], [0, 1, 0], [0, 0, 1]] } }]
}

check_fit

Input:

{
  "scene_id": "ab12cd34ef56",
  "size_m": [2.0, 0.9, 0.8],
  "pose": { "pos": [1.2, 0.4, -0.8], "yaw_rad": 1.57 },
  "clearance_m": 0.3
}

Output:

{
  "ok": false,
  "reasons": ["collides_obstacle:obs_002"]
}

find_free_spaces

Input:

{
  "scene_id": "ab12cd34ef56",
  "size_m": [2.0, 0.9, 0.8],
  "clearance_m": 0.5,
  "yaw_steps": 4,
  "grid_step_m": 0.2,
  "max_results": 3
}

Output:

{
  "candidates": [
    {
      "pos": [4.0, 0.4, 1.2],
      "yaw_rad": 1.57,
      "score": 1.246,
      "notes": ["wall_distance_m=0.184", "center_distance_m=1.430"]
    }
  ]
}

list_scenes

Input:

{}

Output:

[
  {
    "scene_id": "ab12cd34ef56",
    "created_at": "2026-02-18T20:00:00Z",
    "latest_version": 1
  }
]

FastAPI debug endpoints

• POST /scenes/import
  • multipart upload with file field, or JSON body { "path": "/abs/path.glb" }
• GET /scenes/{scene_id}/summary
• GET /scenes/{scene_id}/scene_graph
• GET /health

Data model snapshot

scene_graph.json:

{
  "units": "meters",
  "room": {
    "floor_plane": { "normal": [0.0, 1.0, 0.0], "d": -0.02 },
    "floor_polygon_xz": [[0.0, 0.0], [5.0, 0.0], [5.0, 4.0], [0.0, 4.0]],
    "bounds_aabb": { "min": [0.0, -0.1, 0.0], "max": [5.0, 2.7, 4.0] }
  },
  "obstacles": [
    {
      "id": "obs_001",
      "footprint_xz": [[1.0, 1.0], [1.8, 1.0], [1.8, 1.7], [1.0, 1.7]],
      "obb": {
        "center": [1.4, 0.5, 1.35],
        "extent": [0.8, 1.0, 0.7],
        "rotation": [[1.0, 0.0, 0.0], [0.0, 1.0, 0.0], [0.0, 0.0, 1.0]]
      }
    }
  ]
}

Notes

• Determinism: fixed random seed for mesh surface sampling, deterministic sorting for obstacle IDs and candidate ranking.
• Semantic labels are heuristic and currently include broad classes such as bed, kitchen, table, storage, and unknown.

Web App Workbench

A UI app is included at ./web with:

• source model viewer (GLB/GLTF)
• derived spatial representation viewer (floor + obstacles)
• MCP-backed chat panel with prompt chips

Run it:

cd web
npm install
npm run dev

Then open http://localhost:3000.