mcp3d
A local MCP server for parametric, manufacturing-focused CAD workflows using Build123d as the modeling engine and FastMCP for the protocol.
README
mcp3d
A local MCP server for parametric, manufacturing-focused CAD workflows.
The server uses FastMCP for the Model Context Protocol and Build123d as its modeling engine.
Run
uv run mcp3d
It communicates through standard input/output, so diagnostic logging must always go to standard error.
Model-facing MCP documentation
The runtime documentation is deliberately more complete than this README:
- Server instructions explain the create → inspect → revise → export loop.
- Each MCP tool has a field-level description with valid operation shapes, ordering rules, revision semantics, diagnostics, and repair guidance.
- The
mcp3d://guideMCP resource supplies the complete model guide, including constrained-sketch and feature-graph examples. An agent should read it before it uses an unfamiliar feature type.
The authoritative source for the runtime documentation is
tool_docs.py. This keeps the model-facing guidance
versioned with the actual server implementation.
For new work, prefer the v1 feature graph below. The v0 box/through-hole form remains only for backwards compatibility.
v0 interface
part.applycreates or revises an immutable single-part revision.part.analyzereturns structured checks and PNG orthographic/isometric views.part.exportwrites STEP and/or STL artifacts.
Views use shaded PyVista/VTK rendering with feature edges. If a local machine cannot create an off-screen OpenGL context, the server automatically falls back to its technical SVG projection rasterized as PNG and reports the selected renderer in the tool result.
For known headless environments where VTK may abort before it can raise an
exception, run with MCP3D_RENDERER=technical to force the portable technical
renderer. Normal local use remains PyVista-first.
part.apply returns one isometric image by default. Use
render: {"views":["top","front"]} to request exactly those canonical views,
or render: {"views":[]} when structured checks are sufficient. Use
part.analyze for additional views after the initial revision.
The legacy v0 form supports a box base plus named through_holes features.
All dimensions are millimetres.
Implementation architecture
The MCP adapter is deliberately thin. The internal dependency direction is:
server (FastMCP) → application (revisions) → cad / rendering / artifacts
application/owns immutable revision creation, transactional updates, and in-memory local state.cad/owns Build123d compilation, its private build context, sketches, selectors, datum planes, and feature operations.rendering/owns PyVista, technical SVG fallback, and labelled sketch evidence; it returns neutral in-memory image bytes rather than FastMCP objects or filesystem paths.server.pyis the only production module that creates FastMCPToolResultand image content blocks.
Rendering and sketch evidence are ephemeral: they are encoded directly into
the MCP response and do not create files in .mcp3d. The server creates
.mcp3d/artifacts/<part>/r<revision>/ only when part.export is explicitly
called, and that directory contains the requested STEP and/or STL deliverables.
workspace.py remains as a small backwards-compatible shim for early callers;
new internal code should use PartService and FeatureGraphCompiler directly.
v1.1 feature graph
part.apply also accepts a recipe with named operations. The implemented
workflow is:
box → datum planes → sketches / constraint graphs → solid features
→ finishing, shell/draft, and replicated feature tools
Sketch geometry is authored in its support plane's local 2D coordinates. A
tangent_arc requires two named line guides, a radius, and a deterministic
solution.span ("short" or "long"). Profiles can be polygons or paths
that incorporate named sketch entities. Request a sketch diagnostic with:
{"kind": "render_sketch", "sketch": "relief_sketch"}
Constraint graphs
Alongside direct line and deterministic tangent_arc geometry, a sketch can
contain a constraint_graph. It owns named point, line, and circle
entities; the solver materializes its lines/circles before ordinary sketch
geometry is built. Coordinates are in the sketch plane's local millimetres.
{
"geometry": [
{"id": "p0", "kind": "point", "position": [0, 0]},
{"id": "p1", "kind": "point", "position": [40, 3]},
{"id": "bottom", "kind": "line", "start": "p0", "end": "p1"}
],
"constraints": [
{"id": "origin", "kind": "fixed", "target": "p0"},
{"id": "horizontal", "kind": "horizontal", "target": "bottom"},
{"id": "width", "kind": "distance", "a": "p0", "b": "p1", "value": 40}
]
}
Supported relations are fixed, coincident, horizontal, vertical,
parallel, perpendicular, angle, distance, equal_length, radius,
diameter, equal_radius, midpoint, and branch-explicit
tangent_line_circle. A relation can refer to a point directly, a line's
line.start/line.end, a circle's circle.center, or a projected external
line where appropriate. The initial positions select the local solution
branch. require_fully_constrained: true rejects a profile with remaining
local degrees of freedom.
The revision report exposes solver status, local DOF/Jacobian rank, residuals
per constraint, locally redundant constraints, and suspected conflict IDs.
render_sketch returns a labelled image with entity/point names, dimensions,
and solver status; it is the intended visual review loop before export.
Solid operations
The feature graph supports:
- Base and reference:
box,datum_plane(explicit,offset_from_face,rotate_about),sketch - Profile features:
extrude,revolve,loft,sweep - Finishing:
fillet,chamfer,shell,draft - Feature-tool replication:
linear_pattern,polar_pattern
extrude, revolve, loft, and sweep generate named feature tools with
an operation of "add" or "cut". Patterns copy that tool—not the whole
current part—and their count includes the original source. sweep.path is
"sketch_id.entity_id"; loft.sections is a list of closed-profile sketch
ids.
Finishing operations use geometric queries evaluated at the exact operation, never B-rep edge indexes. A selector requires an expected cardinality, for example the four outer top edges of a box:
{
"geom_type": "line",
"at_extreme": {"axis": "z", "which": "max"},
"expect": 4
}
Selectors also support parallel_to for edges, normal_to for faces, and
at_extreme with mode: "touches" when containment in the extreme plane is
not required. A mismatch fails explicitly as SELECTOR_EMPTY or
SELECTOR_AMBIGUOUS rather than silently choosing a different edge.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。