Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@brickbuilder-mcp Build this photo as a LEGO model at minifig scale, then export it"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Show Claude a photo, get a buildable LEGO model back.
An MCP server that lets Claude design LEGO models brick by brick, usually from a
reference picture, check that they actually hold together, and export them as LDraw
.ldr files that Mecabricks and
BrickLink Studio open directly.
Every build below was made by Claude through this server from the photo on the left. Nothing was placed by hand.
The models are real, buildable LEGO: standard parts in real colours, bricks laid in staggered courses, and a connectivity check that flags anything floating before export.
Related MCP server: SolidWorks MCP Server
Needs Python 3.11+, uv, curl and unzip.
git clone https://github.com/ElRashoMacuin24/brickbuilder-mcp.git
cd brickbuilder-mcp
scripts/fetch_ldraw.sh # downloads the LDraw parts library (~145 MB zip) into data/ldraw
uv sync
claude mcp add brickbuilder -s user -- uv --directory "$PWD" run brickbuilder-mcpRestart Claude Code, then run /mcp and check that brickbuilder is listed.
Other MCP clients work too: run uv --directory /path/to/brickbuilder-mcp run brickbuilder-mcp over stdio.
Paste or drag an image into Claude Code (or give a path) and say what you want. The more you say about scale, the better:
Build this scene in LEGO without the people. Make it minifig scale, so the chairs are just big enough for a minifig to sit in.
Build this house in LEGO, about 16 studs wide: ~/Pictures/house.jpg
Make a 48×48 flat mosaic of this logo: ~/Pictures/logo.png
Useful things to specify:
Scale: "minifig scale" (doors, seats and stairs sized for minifigures), "microscale", or a size in studs ("about 32 studs wide").
What to leave out: people, background, text.
Colours: "use only common colours" keeps it cheap to build for real.
Claude plans the scale, builds bottom-up, renders previews next to your photo and fixes what doesn't match. Ask for changes in plain language:
The orange is too brown, use a redder shade. Make the tower taller and add windows. The left rack is hidden behind the bench, move it.
When you're happy, ask Claude to export. The file lands in exports/<name>.ldr.
Mecabricks: Workshop → File → Import → LDraw, then pick the file. You can publish it to the Mecabricks gallery from there.
BrickLink Studio: File → Import → Import LDraw. Studio also gives you a parts list and prices to buy the bricks. It handles very large models better than a browser tab.
If Mecabricks shows missing parts, ask Claude to re-export with legacy_ids (older part
numbers such as 3023 instead of 3023b).
Models autosave to models/<name>.json. Say "open the bowling alley model" in a later
session to keep editing. Copy models/ and exports/ to another machine to take your
builds with you.
Tool | Purpose |
| Models autosave to |
| Curated catalog, full LDraw library search, colour codes |
| Place parts on the stud grid, with collision checks |
| Colour layer maps packed automatically into bricks/plates/tiles in running bond |
| Quantize pictures to LEGO colours; build mosaics |
| Per-level maps, part counts, floating/loose-part detection |
| Preview sheet (iso/front/top/etc.), optionally beside the reference picture |
| LDraw file, one build step per level; |
x: studs to the right; z: studs toward the back (z=0 is the front); y: plates up (brick = 3).
A part's position is the min corner of its body after rotation.
rot 0/90/180/270 (or front/right/back/left). A slope's sloped face points
that way. At rot 0 a 2x4 brick is 4 studs long in x.
Any LDraw part can be placed. Parts outside the curated catalog are handled as their bounding box for collisions, connectivity and previews. The exported file always uses the real part.
Previews are simplified geometry: round parts and arches show as boxes.
Placing tens of thousands of parts one tool call at a time doesn't scale, so for
Helm's Deep (192 × 160 studs, 41,864 parts) Claude wrote a generator instead.
scripts/helms_deep/ contains it:
gen.py designs the scene as a colour voxel grid (terrain, walls, stairs, towers) and
hollows it into a supported shell.
slopes.py adds 45° slopes along the rock ledges.
pack.py packs everything into staggered bricks and plates, then repairs any floating groups.
uv run python scripts/helms_deep/gen.py
uv run python scripts/helms_deep/slopes.py
uv run python scripts/helms_deep/pack.pyThen ask Claude to open_model helms_deep and export it.
scripts/showcase/ has a GPU renderer (OpenGL via moderngl) for nicer images than the
built-in previews, plus the scripts that made the pictures on this page:
uv run --with moderngl python scripts/showcase/glrender.py helms_deep out.png # one image
uv run --with moderngl python scripts/showcase/compare.py # photo vs build
uv run --with moderngl python scripts/showcase/video.py # 9:16 build videocompare.py and video.py read reference photos from scripts/showcase/refs/. That
folder isn't in the repo, so put your own photos there. video.py needs ffmpeg.
The server runs locally over stdio and makes no network requests (only
scripts/fetch_ldraw.sh downloads, from library.ldraw.org). Tool inputs are
treated as untrusted:
Part ids must be plain LDraw names; library lookups can't leave data/ldraw.
Image tools only open picture files (.png, .jpg, …) and refuse decompression bombs.
export_ldr only writes .ldr / .mpd files; models are saved under models/
with sanitised names.
Render size, grid size and parts per call are capped so one call can't exhaust memory.
It does read picture paths and write .ldr files wherever the MCP client asks, with
your user's permissions, so review tool calls as you would any other file access.
Code: MIT (see LICENSE). Parts geometry comes from the
LDraw parts library, which is downloaded separately and is
licensed CC BY 2.0 by its authors; it is not included in this repository.
The reference stills shown next to the builds are from The Phoenician Scheme (2025), The Big Lebowski (1998) and The Lord of the Rings: The Two Towers (2002). They are shown small, for comparison only, and remain the property of their owners.
LEGO® is a trademark of the LEGO Group, which does not sponsor or endorse this project.
Place one or more parts. Parts that collide or are invalid are skipped and reported; the rest are placed. Returns the new part ids.
Inputs: parts (required). Returns: result
Build a picture directly as a mosaic in the current model. upright: a wall facing the front, 1 stud deep (kind=plate gives fine vertical detail, kind=brick is sturdier). flat: lying on the ground, one plate thick, image top = back.
Inputs: width (required), image_path (required), x, y, +6 more. Returns: result
Check structure: floating groups (not connected to anything touching the ground) and loose groups (standing on the ground but not attached to the main body).
Inputs: no named parameters. Returns: result
Summarise the model: bounds, part counts, and top-down colour maps of plate levels
(first row = back, last = front; uppercase letters = colours in the legend).
Maps are shown for levels, or every level when all_levels is set.
Inputs: levels, all_levels, list_all_parts. Returns: result
Write the model as an LDraw .ldr file (one build step per level). Import it in Mecabricks via Workshop > File > Import. If some parts come in missing, re-export with legacy_ids=True (uses e.g. 3023 instead of 3023b).
Inputs: path, legacy_ids. Returns: result
Build volume from colour maps, automatically packed into bricks/plates/tiles.
layers: bottom-to-top list of layers; each layer is a list of equal-length rows seen from ABOVE, first row = back, last row = front. Each character is a cell of 1 stud; '.' = empty. legend maps characters to colours, e.g. {"R": "Red", "W": 15}. Each layer is one brick (3 plates) or one plate tall depending on kind; layer i sits at y + i*height. (x, z) is the front-left corner of the map.
upright=True: each layer is instead a front-facing wall picture (first row = top, last row = bottom), one course per row, at depth z + layer index. Handy for mosaics and facades.
Cells already occupied are skipped. Seams are staggered between courses.
Inputs: layers (required), legend (required), x, y, +3 more. Returns: result
Downsample a picture into rows of colour letters using LEGO colours, ready for fill_layers. crop = [left, top, right, bottom] fractions (0-1) to focus on the subject. cell sets the cell's aspect ratio when height is omitted: 'stud' for top-down maps, 'brick' or 'plate' for upright walls. Transparent pixels become '.'.
Inputs: width (required), image_path (required), cell, crop, +3 more. Returns: result
List LDraw colours (code, name, hex). Solid colours only unless include_special.
Inputs: search, include_special. Returns: result
List saved models.
Inputs: no named parameters. Returns: result
List the curated parts (brick, plate, tile, slope, round, misc) with their size, or search the whole LDraw library by words in the part title (e.g. 'window 1 x 2'). Size is studs along x x studs along z at rot 0, and height in plates.
Inputs: search, category. Returns: result
Start a new empty model (replaces the open one; models autosave).
Inputs: name (required), description. Returns: result
Open a previously saved model by name.
Inputs: name (required). Returns: result
Remove parts by id and/or every part touching an inclusive grid region.
Inputs: ids, region. Returns: result
Render a preview sheet. views: any of front, back, left, right, top, iso_front_left, iso_front_right, iso_back_left, iso_back_right, or 'azimuth,elevation' in degrees. Default: iso_front_left, iso_back_right, front, top. reference_image adds the source picture as the first panel for side-by-side comparison. Parts are simplified (round parts and non-catalog parts show as boxes).
Inputs: size, views, reference_image.
Undo the last building action.
Inputs: no named parameters. Returns: result
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 15 tools
Each tool targets a clearly distinct action and resource: model lifecycle (open/list/new), part/color discovery, placement/removal, procedural fills, image conversion, structural checks, description, rendering, and export. The only potential overlap between fill_layers and build_mosaic is resolved by their descriptions, which distinguish volume construction from direct picture mosaics.
Most names follow a predictable snake_case verb_noun pattern (open_model, list_parts, add_parts, remove_parts, check_model, render_model, export_ldr). Minor deviations include undo (single verb) and image_to_grid (noun-to-noun), but overall the set is readable and consistent.
Fifteen tools is well-scoped for a LEGO modeling server, covering discovery, construction, validation, rendering, and export without obvious redundancy. Each tool appears to earn its place in the generation workflow.
The surface covers model creation, part placement/removal, color lookup, procedural fills, image-to-grid conversion, structural checking, rendering, and LDraw export. Minor gaps exist: no direct move/rotate/edit operation for existing parts, no delete_model, and no redo, but agents can work around these by removing and re-adding parts or starting a new model.
Design custom LEGO brick sets with 3D models, instructions, kit prices and customer-paid checkout.
CAD in ChatGPT and Claude: editable parts, STEP/STL/drawings, in-chat 3D, automatic checks.
Make any photo or video you can describe, right from Claude.
Turn Claude into a creative studio: DNA-locked characters, images, video, voiceover — 55 tools.
