Agent-native floor plan studio. Humans draw, agents measure — over WebMCP.
Groundplan demonstrates solid tool design with clear naming conventions (verb_noun pattern), comprehensive parameter descriptions, and well-structured schemas. All 20 tools have descriptions (avg ~150 chars) and input schemas with typed parameters. Tool names are action-oriented (get_, check_, read_, add_, edit_, delete_, analyse_, apply_, export_, share_). However, output schemas are not explicitly documented in the visible code, responses are described narratively rather than as formal JSON Schema. Error handling includes hints and recovery guidance (e.g., 'hint' fields in replyError calls), but not all tools document error categories or retryability. Parameter descriptions are generally strong (avg ~80 chars) with enums for constrained inputs (room types, sides, opening kinds). The mutate() wrapper shows thoughtful composition with approval gates and issue tracking. Missing: explicit output schema documentation, per-tool error classification, and some parameter constraints (e.g., numeric ranges for mm values).
Place a piece of furniture in a room. Returns the furniture id and the issues the change introduced or resolved.
Add a door, window or archway to a room. Returns the opening id and the issues the change introduced or resolved.
Add a new room to the plan. Rooms are axis-aligned rectangles that meet edge to edge. Returns the room id and the issues the change introduced or resolved.
Trace the widest route from the entrance to a room, showing where a body of a given width would get stuck. Returns a journey with points, clearance at each point, and the distance travelled.
Apply the standard repair for a violation. Returns the issues the change introduced or resolved.
Read the current plan and the rules it is measured against. Returns a list of violations, each with a suggested fix.
Output schemas not formally documented. Tool descriptions state what is returned (e.g., 'Returns the room id and the issues...') but no explicit JSON Schema for response objects. LLMs cannot reliably parse unstructured responses or plan downstream tool calls without knowing field names and types.
Numeric parameters lack explicit range constraints. Parameters like 'body_diameter_mm', 'width_mm', 'depth_mm', 'offset_mm', 'rotation_deg' have no documented min/max values. LLMs may pass absurd values (negative dimensions, 360+ rotations) that break the tool.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-23 | C | 60 | 2026-07-28+ | v2 |
Remove a piece of furniture. Returns the issues the change introduced or resolved.
Remove a door, window or archway. Returns the issues the change introduced or resolved.
Remove a room from the plan. Returns the issues the change introduced or resolved.
Move or rotate a piece of furniture. Returns the issues the change introduced or resolved.
Resize, move or change the swing of a door or window. Returns the issues the change introduced or resolved.
Resize or move a room. Returns the issues the change introduced or resolved.
Export a room schedule: name, type, area, and measurements.
Export the plan as an SVG drawing, with or without annotations.
Automatically place furniture in a room based on its type. Returns the issues the change introduced or resolved.
What this page can and cannot do, in one call. Read it before attempting anything unusual, and read it when a request does not obviously map onto a tool — it lists the limits explicitly, so you can tell someone plainly that a thing is not possible here and offer the nearest thing that is, instead of guessing.
Load a built-in sample plan: a starter flat with faults, an accessible bungalow, or an empty canvas.
Read the current floor plan: rooms, openings, furniture, and measurements. Everything the agent needs to understand the drawing.
Give a room a new name. Returns the issues the change introduced or resolved.
Generate a shareable link to this plan. The link encodes the entire drawing in its fragment, so it works without a server.
Error handling lacks categorization. The mutate() wrapper returns errors with hints, but tools do not explicitly classify errors as retryable, user-fixable, or fatal. LLMs cannot determine whether to retry, ask the user, or abandon the operation.
Destructive operations (delete_room, delete_opening, delete_furniture) lack explicit dry-run or confirmation documentation. While the mutate() wrapper supports approval gates, the tool descriptions do not state that confirmation is required or optional.
Parameter 'room' and 'opening' accept flexible identifiers (id, exact name, partial name, type) but descriptions do not explain resolution order or ambiguity handling. If two rooms match a partial name, which is returned? This can cause silent misuse.