FreeCAD MCP
Enables an AI agent to drive a live FreeCAD desktop session, allowing it to create and edit geometry, capture views, and export models through MCP tools.
README
FreeCAD MCP
An MCP server that lets an AI agent drive a live FreeCAD desktop session — the model builds geometry, you watch it appear in the FreeCAD window, and the agent can screenshot the 3D view to check its own work.
Verified against FreeCAD 1.1.3 on macOS.
How it works
MCP client (opencode)
│ stdio / MCP
▼
freecad-mcp ← own uv venv, Python 3.10+, mcp SDK
│ XML-RPC on 127.0.0.1:9875
▼
FreeCADMCP addon ← runs inside FreeCAD's bundled Python 3.11
│ QTimer task queue
▼
FreeCAD GUI main thread ← documents, geometry, 3D view
Two processes, deliberately. The MCP server never imports FreeCAD, so it can use a modern venv and any dependencies it likes, while the addon stays stdlib-only inside FreeCAD's interpreter.
The RPC server runs on a background thread, but every FreeCAD call is marshalled
onto the Qt main thread through a queue drained by a QTimer
(addon/FreeCADMCP/freecad_mcp_bridge/dispatcher.py). Touching FreeCAD documents from
a worker thread corrupts state or crashes the app, so this hop is not optional.
The wire protocol is a single call(method, payload_json) -> json endpoint. Passing
JSON through one method sidesteps every XML-RPC marshalling quirk (no None, no
nested heterogeneous types) and keeps errors as structured data instead of faults.
Setup
make install # create the venv, install dependencies
make addon-install # symlink the addon into FreeCAD's Mod directory
make freecad # start FreeCAD -- the bridge autostarts
make check # confirm the bridge answers
make check should print your FreeCAD version and open documents.
The addon is symlinked, not copied, so edits in this repo take effect on the next
FreeCAD restart (make freecad-restart).
Wiring it into opencode
opencode.json already registers the server. Restart opencode to pick it up —
config is read once at startup.
It uses "cwd": ".", which opencode resolves from the workspace directory, so there
is no absolute path to update when you clone this elsewhere.
timeout is raised to 120s because the 5s default is not enough for capture_view
or export_objects on a non-trivial model.
Other MCP clients
Clients without a workspace concept (Claude Desktop and friends) need an explicit
path — replace /absolute/path/to/FreeCAD-MCP with wherever you cloned this:
{
"mcpServers": {
"freecad": {
"command": "uv",
"args": ["run", "--directory", "/absolute/path/to/FreeCAD-MCP", "freecad-mcp"]
}
}
}
Tools
| Tool | Purpose |
|---|---|
freecad_status |
Connectivity check, FreeCAD version, open documents |
list_documents / create_document / open_document |
Document management |
save_document / close_document / recompute |
Persistence and rebuild |
list_objects / get_object |
Inspect the tree; get_object reports volume, area, bounding box |
create_object / edit_object / move_object / delete_object |
Typed geometry editing |
get_selection |
Read what the user selected in the GUI |
run_python / reset_python |
Execute Python inside FreeCAD (escape hatch) |
capture_view |
Screenshot the 3D view, returned as an image |
export_objects |
STEP, IGES, STL, OBJ, PLY, 3MF, AMF, OFF, BREP, FCStd |
Value formats
create_object and edit_object accept JSON and coerce it to FreeCAD's native types:
| Property kind | Accepted |
|---|---|
| lengths | 10, 10.5 (millimetres) |
| vectors | [x, y, z] or {"x":.., "y":.., "z":..} |
| placement | {"Base": [x,y,z], "Rotation": {"Axis": [0,0,1], "Angle": 45}} |
| rotation | 45 (degrees about Z), {"Axis":..,"Angle":..}, {"Yaw":..,"Pitch":..,"Roll":..} |
| links | "Box" — an object Name or Label |
| colours | "#ff8800", [1.0, 0.53, 0.0], or [255, 133, 0] |
Angles are degrees, lengths millimetres.
view_properties targets the ViewObject: ShapeColor, Transparency (0–100),
DisplayMode, Visibility.
Example
Creating a plate with four drilled holes:
create_object type_id="Part::Box" properties={"Length":60,"Width":40,"Height":10}
create_object type_id="Part::Cylinder" properties={"Radius":3,"Height":20,
"Placement":{"Base":[10,10,-5]}}
create_object type_id="Part::Cut" properties={"Base":"Box","Tool":"Cylinder"}
capture_view # look at it before declaring success
export_objects path="~/plate.step"
Configuration
| Variable | Default | Applies to |
|---|---|---|
FREECAD_MCP_HOST |
127.0.0.1 |
both |
FREECAD_MCP_PORT |
9875 |
both |
FREECAD_MCP_TIMEOUT |
300 (seconds) |
MCP server |
FREECAD_MCP_AUTOSTART |
1 |
addon |
The addon also reads FreeCAD parameters under
BaseApp/Preferences/Mod/FreeCADMCP (Host, Port, AutoStart), editable via
Tools > Edit parameters. Environment variables win.
The bridge binds to loopback only and has no authentication — anything that can reach
the port can execute arbitrary Python in FreeCAD. Do not expose it beyond 127.0.0.1.
Bridge controls in FreeCAD
The addon adds an MCP workbench with Start / Stop / Status commands. The bridge autostarts 1.5s after the GUI is up, and stops cleanly on quit.
Development
make smoke # end-to-end test through a real MCP client (FreeCAD must run)
make lint # ruff
make freecad-restart # reload addon changes
make addon-status # is the addon linked?
make smoke exercises all 18 tools against a live session and verifies boolean
geometry numerically (a drilled plate's volume must match l·w·h − nπr²h).
Two FreeCAD gotchas this codebase works around
1. InitGui.py has no usable module scope. FreeCAD runs it via exec from
inside a function, so top-level names become that function's locals. Any function or
class method defined in InitGui.py resolves its globals against FreeCAD's own init
module and dies with NameError. All logic therefore lives in
freecad_mcp_bridge/workbench.py, which is imported normally; InitGui.py is a
five-line shim.
2. package.xml decides where FreeCAD looks for your code. When a workbench
entry omits <subdirectory>, FreeCAD defaults it to the workbench <name> — so a
workbench named MCP makes FreeCAD search FreeCADMCP/MCP/InitGui.py and silently
skip the addon. Hence the explicit <subdirectory>./</subdirectory>.
A third, smaller one: in FreeCAD 1.x, ShapeColor is a Python shim over
ShapeAppearance rather than a registered property, so getTypeIdOfProperty fails
for it. serialize.py falls back to inferring the type from the value currently
stored.
Layout
addon/FreeCADMCP/ installed into FreeCAD's Mod directory
InitGui.py thin shim (see gotcha 1)
package.xml addon metadata (see gotcha 2)
freecad_mcp_bridge/
workbench.py workbench, commands, autostart
server.py XML-RPC server + JSON envelope
dispatcher.py background thread -> Qt main thread
handlers.py the FreeCAD operations
serialize.py JSON <-> FreeCAD type conversion
src/freecad_mcp/
server.py MCP tool definitions
client.py XML-RPC client
__main__.py CLI entry point
scripts/smoke.py end-to-end test
Troubleshooting
make check says the bridge is unreachable. Confirm FreeCAD is running, the addon
is linked (make addon-status), and check FreeCAD's View > Panels > Report view for
[FreeCAD MCP] messages.
Address already in use on startup. A previous FreeCAD process still holds the
port — make freecad-quit force-kills leftovers.
Addon changes have no effect. The addon is imported once at startup; run
make freecad-restart.
A tool times out. Something is blocking FreeCAD's main thread, usually a modal dialog. Dismiss it in the GUI.
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。