cadkit
An MCP server for parametric part modeling in Onshape, producing fully-defined, variable-driven sketches and features. It enables LLMs to create editable CAD models using semantic selection and geometrically grounded constraints.
README
cadkit — Onshape MCP

Two Model Context Protocol servers for driving Onshape from an LLM:
cadkit— the focus of this fork. A part-modeling server built around one idea: every part should come out idiomatic, fully-defined, and variable-driven — the way a careful human models, not a pile of absolute coordinates. One sketch carries its entities, geometric constraints, and driving dimensions; it's grounded to the origin and parameterized by variables; downstream features select edges and faces by meaning (concave edge, cylindrical face) rather than by transient ids that break when topology shifts.onshape_mcp— the upstream server this repo forked from (hedless/onshape-mcp), kept intact. Broader and assembly-focused (mates, instances, interference, export). Reach for it when you're assembling parts rather than authoring one.
Why two servers?
onshape_mcpis wide;cadkitis opinionated. cadkit deliberately emits a narrower, stricter shape of geometry so that what you get back is parametric and editable, not just present. See PLAN.md for the thesis and roadmap.
Why cadkit models the way it does
A few hard-won principles are baked into the tools (the full list is in PLAN.md):
- Grounded + dimensioned, or it warns.
cad_sketch_closereports whether the sketch is grounded to the origin and dimensioned;require_well_formed=truerefuses to ship an under-defined sketch instead of letting the solver place it unpredictably. - Variables drive geometry, but only where they earn it. A dimension, an extrude depth, a
hole diameter, a fillet radius — all accept a number or an expression /
#variable. Use a variable when a value travels beyond one sketch or is derived (#leg_len - #thick); use a geometric constraint (equal,symmetric) when the relationship lives between entities; use a literal for a true one-off. - Select by meaning, not by id.
cad_find_edges/cad_find_facesreturn deterministic ids chosen by geometry (concave edges for fillets, a cylindrical face by radius), so a feature keeps referring to the right thing after the model changes.
The cadkit tools (34)
| Group | Tools |
|---|---|
| Document / studio | cad_document_create, cad_part_studio_create |
| Sketch session | cad_sketch_begin → cad_sketch_line · cad_sketch_circle · cad_sketch_arc · cad_sketch_fillet · cad_sketch_mirror · cad_sketch_pattern · cad_sketch_rectangle · cad_sketch_polyline · cad_sketch_slot → cad_sketch_constrain · cad_sketch_dimension → cad_sketch_close |
| Variables | cad_set_variable (update-or-create; no duplicates), cad_get_variables |
| Features | cad_extrude, cad_revolve, cad_fillet, cad_chamfer, cad_shell, cad_hole (simple / counterbore / countersink), cad_plane (offset datum plane) |
| Pattern / mirror | cad_mirror, cad_pattern (linear + circular) — feature-based: repeat whole features, not faces |
| Inspection / lifecycle / I/O | cad_measure (count/volume/bbox in one eval), cad_delete_feature, cad_suppress, cad_edit_feature, cad_export (STL/STEP/…), cad_api_calls (running quota counter) |
| Semantic selection | cad_find_edges (circular / concave / convex / linear / extreme / on-plane), cad_find_faces (planar-by-normal / cylindrical / largest / smallest / extreme / adjacent-to-extreme / on-plane) |
A sketch is one session: cad_sketch_begin(plane=… | face=<id>), add entities, add geometric
constraints + driving dimensions, then cad_sketch_close. cad_pattern / cad_mirror take the
featureIds of the features to repeat (e.g. an extrude or hole) plus a direction/axis/plane.
A small example (L-bracket, parametric)
cad_set_variable("leg", "2 in"); cad_set_variable("thick", "0.25 in")
sk = cad_sketch_begin(plane="Front")
# draw the L profile with cad_sketch_polyline / line, then:
cad_sketch_constrain("ground_origin", a="ln1.start") # pin to origin
cad_sketch_dimension(kind="length", target="ln1", value="#leg")
# ... fully define ...
cad_sketch_close(require_well_formed=true) # refuses if under-defined
cad_extrude(sketchFeatureId=sk, depth="#thick", operation="NEW")
Install
Both servers live in this one repo and share the same Onshape client.
git clone https://github.com/saltyeg/cadkit-mcp.git
cd cadkit-mcp
python -m venv venv && source venv/bin/activate
pip install -e .
Credentials
You need an Onshape API access key + secret key from the Onshape Developer Portal. cadkit resolves them in this order:
ONSHAPE_ACCESS_KEY/ONSHAPE_SECRET_KEYenvironment variables, else- the
onshapeMCP server entry already in your~/.claude.json(so if you've configured theonshape_mcpserver, cadkit reuses those keys — you never paste them twice).
See .env.example for all supported variables.
OAuth2 (bring your own account)
cadkit also supports OAuth2 — authenticating as the user via their own Onshape account instead of API keys. This is the architecture behind the "bring-your-own-agent" model and the path to an App Store listing whose calls are exempt from the per-user annual quota. It's optional; API keys remain the default and nothing changes if you don't use it.
- Register a Connected desktop app (OAuth application) in the
Onshape Developer Portal with:
- Redirect URI
http://localhost:8910/callback(must match exactly) - Scopes read, write, delete (
OAuth2Read OAuth2Write OAuth2Delete)
- Redirect URI
- Put the issued credentials in
.env(or your environment):ONSHAPE_OAUTH_CLIENT_ID=… ONSHAPE_OAUTH_CLIENT_SECRET=… - Run the one-time browser handshake:
cadkit-auth login # opens the browser, stores tokens at ~/.cadkit/onshape_token.json cadkit-auth status # show token + expiry; cadkit-auth logout to revoke locally
Once a token is stored, the cadkit server prefers it over API keys automatically (with silent refresh). Caveat: until the app is published in the App Store it's a private OAuth app, so calls still count against the 2,500/yr quota exactly like API keys — OAuth buys the right architecture, not quota relief, until launch.
Register with Claude Code
Add to ~/.claude.json (or ~/.claude/mcp.json). The cadkit entry can omit keys if the
onshape entry already has them:
{
"mcpServers": {
"onshape": {
"command": "/abs/path/to/onshape-mcp/venv/bin/python",
"args": ["-m", "onshape_mcp.server"],
"env": { "ONSHAPE_ACCESS_KEY": "…", "ONSHAPE_SECRET_KEY": "…" }
},
"cadkit": {
"command": "/abs/path/to/onshape-mcp/venv/bin/python",
"args": ["-m", "cadkit_mcp.server"]
}
}
}
Use the absolute path to the venv's Python. Restart Claude Code after editing the config or after any change to the server code (a running server holds the old module until restart).
API quota — read this before testing
The Onshape free / standard / education-student tiers allow 2,500 successful API calls
per user per year (limits doc). Only
2xx/3xx responses count — 4xx/5xx are free; 429 is a separate short-term burst limit;
402 means the annual budget is spent. A naive live test suite drains this fast.
cadkit's development practice is built around that constraint:
- Offline-first tests.
tests/test_cadkit_builders.pyasserts on the JSON the builders emit — parameter ids, enum strings, constraint structure — with zero API calls. This catches the bug class that actually costs debugging time (a wrong/hiddenparameterId). cadkit_mcp/devkit.py— quota-frugal live helpers:ScratchStudioreuses one part studio across checks;measure_fsreturns bbox / volume / count / sketch bbox / named variables in a single FeatureScript eval (SOLID-body filtered, inch-converted).- One on-demand live smoke (~6–8 calls), run by hand before a release — only for truths offline can't prove (a variable actually drives geometry; a concave edge actually fillets).
# offline builder tests — free, run anytime
venv/bin/python -m pytest tests/test_cadkit_builders.py -o addopts="" -q
Layout
cadkit_mcp/
├── server.py # the cadkit tools + feature JSON builders
├── sketch.py # SketchSession: entities, constraints, grounding, diagnostics
├── selection.py # semantic edge/face finders (FeatureScript-backed)
├── quota.py # successful-call counter behind cad_api_calls
├── hole_template.json # 160-param native Hole feature template
└── devkit.py # quota-frugal live verification helpers
onshape_mcp/ # inherited upstream server (assemblies, mates, export) — see git history
tests/ # offline builder tests (zero-API)
PLAN.md # cadkit thesis + roadmap (P0–P3)
Roadmap
cadkit is ordered so the parametric core is correct before feature breadth — see
PLAN.md. Shipped (P0–P2): the sketch session, idempotent variables, parametric
scalars everywhere, extrude/revolve/fillet/chamfer/shell/hole, sketch-on-face, feature-based
pattern/mirror, semantic edge/face selection, plus inspection/lifecycle/IO — cad_measure,
cad_get_variables, cad_delete_feature, cad_suppress, cad_edit_feature, cad_export.
In progress (P3): richer semantic selection (largest/smallest/extreme/adjacency/on-plane done;
by-tag next) and sketch ergonomics (slots, center-point arcs, fillets, construction geometry
done; in-sketch mirror live-verified; in-sketch pattern emits geometric copies). Deferred: cad_rollback, mass/COM in
cad_measure, and tapped threads on the native Hole feature.
Notes
Not tested on assemblies yet, only individual parts
Credits
Forked from hedless/onshape-mcp; the onshape_mcp
server and its assembly tooling are upstream's work. Built on the
Model Context Protocol and the
Onshape REST API. MIT License.
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。