## Laydyne (spatial modelling over MCP)

Use Laydyne's MCP tools when the task is to model, check, compare or hand over a space (floors, buildings, layouts, processes). You edit the Laydyne Studio tab the user has open.

Connect your AI to Laydyne over MCP with OAuth: add the server https://laydyne.com/api/mcp/oauth and sign in with a free Laydyne account when asked. Keep Laydyne Studio open in your browser; tool calls run in that tab. Codex: `codex mcp add laydyne-oauth --url https://laydyne.com/api/mcp/oauth` then `codex mcp login laydyne-oauth`. Codex asks before each Laydyne edit; to let it edit without asking, add `default_tools_approval_mode = "approve"` under `[mcp_servers.laydyne-oauth]` in ~/.codex/config.toml. Claude Code: `claude mcp add --transport http laydyne-oauth https://laydyne.com/api/mcp/oauth` then `claude mcp login laydyne-oauth` (or /mcp inside Claude Code). Other apps and the agent-key alternative: https://laydyne.com/docs#connect-mcp.

### Rules

- Laydyne edits the Studio tab the user has open: the work lives in that tab. Save it (save_local_project) or back it up (export_project) before you finish, and before any long pause.
- Call Laydyne tools one after another, not in parallel: the Studio tab works through a few calls at a time and answers the rest with code busy.
- Start with get_space {detail: "summary"} for the revision and the floor ids. Pass expectedRevision to every change, and take the next one from each result’s revision instead of reading get_space again. Ask create_space and edit_space for result: "brief" when you do not need every part back.
- A result’s JSON is its last text block (there is no structuredContent): use the fields you need rather than printing whole results. Part sizes, capabilities and the coordinate convention come once, from get_space {detail: "about"}.
- Look at what you built: get_view_image draws the 2D plan or a 3D view without changing the user’s screen. Do not report a layout you have not looked at.
- Run check_space (floors: "all" for a building) after changes and fix what it finds: overlaps, passages that are too narrow, doors that open onto nothing. A finding you keep on purpose is no reason to stop: say which and why, then go on saving and comparing (a saved option keeps its findings).
- Place parts by kind (get_part_catalog lists them with sizes); use addArray for repeated shelves, desks or seats. Machines, conveyors, pallet racks, workbenches, sofas, kitchen counters and washing machines take a variant (a CNC lathe, a curve section, an L-shaped sofa, a counter with a sink or a hob): get_part_catalog {id: "machine"} says how to tell each from a photo. Give parts placed from evidence a source (drawing, measured or assumed) and a short note.
- For large or rule-based models, generate them with the kit (get_project_schema part: "overview", /sdk/laydyne-kit.mjs) and send them with apply_changes: dryRun first, then apply.
- Drawings go in with set_underlay (PDF pages too) and trace_underlay; check the fit with measure_underlay_fit. Do not process the picture pixel by pixel yourself.
- Homes: give the space kind home (room doors from 0.75 m pass the check; entrances keep 0.9 m); get_part_catalog {use: "home"} lists the bathtub, TV board, shoe cabinet and washing machine. A balcony people walk on is floor: draw it into the outline. Eaves and canopies that reach past the floor need room first: edit_space room.overhang (any side; the floor and everything on it stay where they are, coordinates then start further west and north).
- Sizes read from photos are estimates until one is measured: one measured length fixes the whole room with rescale_space (parts keep their sizes and move with it; it also checks doors, pallets and beds against the scale). Say what is estimated in the part notes and in your answer.
- Where a photo was taken from is solved, not guessed: solve_photo_camera from 6 or more points of the room itself (corners at floor and ceiling, column bases, door and window frames; never the equipment you are placing; columns counted from the far wall), then get_photo_overlay shows the model over the photo. Give fov or focal35mm when the photo’s EXIF has it.
- Processes: read get_process_catalog before writing a scenario. A job stays where it arrived until a goto moves it; give a line’s work an item agent and goto the line’s output after the equipment is released. A result is stale when its spaceHash differs from the building’s: run it again before quoting it.
- Compare options with save_design_option / compare_design_options (layout) and save_process_option / compare_process_options (process, same arrivals and seed). Report differences inside the noise as no difference.
- To find the best placement or fleet rather than compare two you made, use search_layouts: decide what may change (parts along a line or in an area, turns, spacing, kept stations, scenario parameters), the objective (one existing measure) and the constraints (no new check findings, minClearance, keep-outs, fixed parts, thresholds); search small first, read the evidence (intervals, withinNoise, rejections), then widen it, apply the best with its edits only when the user asks, or compare its saved option. Call the result the best of the candidates searched, not a proven optimum.
- Cameras: use suggest_cameras instead of placing them by eye. Give what must be seen (targets: areas, entrances or tills with around metres, weights, maxDistance or pixelsPerMetre for detail), where they may go (walls, columns, a ceiling grid or points; keep-outs), the lens and how many; report coverage per target, blind spots and upperBound, and call it a greedy suggestion (at least 63 % of the best the same number of candidates reach; not a proven optimum). Place them with its apply input only when the user agrees; camera parts (kind camera, their view in sight) are what get_visibility checks after any edit.
- Hand-over: get_view_image with delivery: true gives the paper sheet with dimensions; format: "pdf" with planLayout: "sheets" gives every level as a PDF page, in base64 pieces with nextOffset and sha256. To issue drawings, call issue_drawings (revision, status, recipients): it records the issue in the project’s register, numbers the sheets, fills the revision table, clouds what changed since the previous issue and returns the PDF’s first piece (the rest with get_view_image {issue, format: "pdf", offset}); get_issue_register lists what was issued (csv to keep). Keep revision words out of floor, part and project names, which every later sheet repeats.
- Say what Laydyne does not calculate when it matters to the answer: collisions and avoidance, crowd physics, evacuation, real lift dispatch, revenue. Synthetic examples are not measurements of a real site.

### Photos: quick, careful or detailed

Users choose how carefully a room is built from photos — quick, careful or detailed; ask when they have not said.

- Quick: the room's shape from the photos, one measured length applied with rescale_space (one call), the furniture and equipment placed by eye. No camera solving: one call more than building by eye. The room's size is as good as the measured length; positions placed by eye stay as rough as the eye.
- Careful: also solve_photo_camera on the 3 or 4 photos that show the room's corners, places parts at the floor points its locate gives, and checks each solved photo with get_photo_overlay. Each solved photo adds 2 or 3 calls (one solve, the overlay, a second piece when the photo is over 200,000 base64 characters); a solve answers in about 1-2 KB, an overlay with a 1024 px picture. In the factory test, 10-16 points read by eye put the cameras 0.05-0.72 m from where the photos were taken (median 0.33 m; 0.24 m when the lens was given). Reading the points means looking closely at each photo: in that test, 3-6 close-up views of each photo on the agent's side.
- Points for a camera: the room itself, at different depths and heights, spread over the picture. Equipment is what you are placing: its corners make a camera fit a wrong model. A generated or retouched photo draws column tops and frames loosely; prefer where things meet the floor and window and door frames.
- solve_photo_camera says when a point disagrees (it sets a few aside when most agree) and when the room would fit deeper or taller; its uncertainty says when the points do not pin the camera down (points all on one wall, a facade seen straight on). get_photo_overlay needs the photo's file in the tool call: agents that run code can send it; others can still solve cameras and use locate.
- Detailed (after quick or careful, on the same model, no new tools): finer parts and variants (get_part_catalog lists them), colours and small items such as door frames and fittings, and the photos' misfits narrowed. Build quick first and refine; do not rebuild.
- Display levels are the user's choice in View settings, or set_space_view display: light (for devices without a graphics processor), standard, path-traced (while the view is still, the browser computes light and reflections on the same shapes). For a picture with computed light, get_view_image pathTrace: true (128 samples or 15 s; pass {samples, seconds} for more, up to 20 s). It is a display only: sizes, quantities and calculations do not change.

### Starting prompts

- Build a first space: A small shop from a one-line brief, checked and shown in 2D and 3D.
- Build from a floor plan: Lay a drawing (PDF or image) under the floor, trace it and build the model from it.
- Model a room from photos (quick): The room’s shape and size from photos and one measured length, the furniture or equipment by eye. No camera solving: one call more than building by eye.
- Model a room from photos (careful): Also solves where each photo was taken from and checks the model over the photos. In the factory test the solved cameras were 0.05–0.72 m from where the photos were taken (median 0.33 m); 2–3 more calls a photo.
- Change a layout and keep both versions: Save the current layout as an option, make a change, then compare and keep both.
- Compare layouts under the same demand: Set up a process (arrivals, work, resources), run it on two layouts and compare.
- Search for the best layout: Let Laydyne build and simulate the placements and counts you allow, and rank them with evidence.
- Suggest where to put cameras: Security or streaming cameras placed so the floor, the entrances or a stage are seen, with coverage, blind spots and how much better any placement could be at most.
- Hand over drawings and quantities: Scaled drawings as PDF, the quantities and a project file to keep.

The full prompts are in https://laydyne.com/llms-full.txt and, once connected, in the server's MCP prompts.