autodesk-inventor-mcp
MCP server for Autodesk Inventor that lets AI clients model parts, build assemblies, and inspect models via COM automation.
README
autodesk-inventor-mcp
An MCP server for Autodesk Inventor. It lets an AI client — Claude, Cursor, or anything else that speaks MCP — model parts, build assemblies, and inspect what it just made.
Autodesk ships official MCP servers for Fusion, Revit and Product Help. Inventor has none, and none is announced. This fills that gap.
Status: alpha. Developed and tested against Inventor 2027 on Windows 11.
Why it works differently from the Fusion server
Fusion's MCP server runs inside Fusion, because Fusion has no out-of-process API — an add-in is the only way in. Inventor exposes a full COM automation interface, so this server is an ordinary external Python process:
- nothing to install into Inventor, and no add-in to rebuild per version
- a crash in the server does not take Inventor with it
- normal Python tooling and debugging
Things that turned out not to be true
Several widely repeated claims about Inventor automation did not survive
measurement, and each one changed the design. All of them have a test that would
fail if they stopped holding — details in docs/design.md §2.
- Out-of-process COM does produce real solids (
IsSolid == True, and a cylinder with 3 faces). The folklore that only in-process execution — an add-in or an iLogic rule — yields solids does not hold on 2027. If you got surfaces, look at how the profile was created, not at the process boundary. iLogicAutomation.AddRuleexecutes the rule as it adds it. Code that callsAddRuleand thenRunRulebuilds everything twice, and a trailing boolean usually hides the evidence.- A
gen_pycache does not break.ComponentDefinition— the usual telling of the late-binding rule. What it actually does is hand back typed wrappers for child objects that expose only their declared interface, so a STEP translator arrives with noHasSaveCopyAsOptions, a feature with no.Operation, and an occurrence definition with no.WorkPlanes. - Inventor's object names are localised. On a zh-CN install the origin
planes are
'XY 平面'and the structured BOM view is'结构化'. Any code that looks up"XY Plane"or"Structured"is en-US only. This server addresses them by index behind a language-independent shorthand. - Patterned instances are not new bodies. A circular pattern adds lumps to the seed's body; the body count does not move.
Install
Requires Windows and a licensed Inventor installation.
git clone https://github.com/xtylerai2026-oss/autodesk-inventor-mcp
cd autodesk-inventor-mcp
uv venv
uv pip install -e .
Claude Desktop / Claude Code
{
"mcpServers": {
"inventor": {
"command": "path/to/autodesk-inventor-mcp/.venv/Scripts/inventor-mcp.exe"
}
}
}
HTTP transport
inventor-mcp --http --port 27183
Then point the client at http://127.0.0.1:27183/mcp. Running over HTTP keeps
the Inventor connection warm across client restarts and can be poked with
curl.
Usage
Start Inventor (or let the server start it), then just ask:
Make a 60 mm diameter, 20 mm tall boss with a 12 mm bore through it, and tell me the mass in aluminium.
The interesting part is not that it builds — it is that it checks.
describe_model returns is_solid, face counts, volumes, and per-feature
health, so a client can notice it accidentally produced a surface, or that a
fillet failed, and fix it without being told.
Tools
24 tools in six groups. Full signatures in docs/design.md §5.
| group | tools |
|---|---|
| Session | inventor_status new_document open_document save_document close_document |
| Perception | describe_model measure screenshot get_parameters set_parameters |
| Part modelling | sketch_and_extrude loft_sections revolve_profile pattern_bodies boolean_combine |
| Assembly | place_occurrence pattern_occurrences add_constraint check_interference get_bom |
| Import / export | import_geometry export_document |
| Escape hatches | run_python_com manage_ilogic_rule |
Read-only tools carry readOnlyHint, so clients can auto-approve inspection
calls. That matters more than it sounds: the build-inspect-adjust loop only
works if inspecting is free.
Units
Every length is millimetres and every angle is degrees, at every boundary. Inventor's own API works in centimetres and radians; this server converts at the edge so you never meet a silent 10× error.
Scope
Supported: part documents (.ipt) and assembly documents (.iam).
Assembly constraints work against named entities — work planes, axes, points,
iMates. Constraints that require picking a particular face or edge are out of
scope on purpose: Inventor has no stable name for a face and topological ids
move on rebuild. Use run_python_com when you genuinely need one.
Not supported: drawings (.idw), sheet metal, CAM.
Development
uv pip install -e ".[dev]"
pytest # 24 unit tests, no Inventor needed
pytest -m inventor # 32 integration tests, drives a real Inventor
Current state on Inventor 2027 (build 310192060): 24 unit + 32 integration tests passing, 0 Inventor processes left behind.
Integration tests attach to a running Inventor if there is one and launch one
otherwise — and call Quit() on teardown only if they launched it. Please
keep it that way. Automation that forgets is how you end up with a pile of
headless Inventor processes holding gigabytes; this project started by cleaning
up nine of them.
Licence
MIT
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。