Ged Invest MCP
An MCP server for construction tools that calculates wall formwork material takeoffs from geometry inputs, enabling LLMs to generate accurate bills of materials.
README
Ged Invest MCP
An MCP (Model Context Protocol) server hosting a growing set of construction tools for Ged Invest. The first tool domain is wall formwork quantity takeoff; more tool domains can be added later as separate submodules.
Idea / architecture
Separation of concerns:
Photo / drawing / PDF
│
▼
[ LLM (ChatGPT / Claude) ] ← reads the geometry, builds structured data
│ { segments:[{length, height, corner}], system:"BAUTEKK" }
▼
[ Ged Invest MCP (Python) ] ← deterministic calculators (tools)
│ panel selection + corners + hardware + aggregation
▼
[ Result: BOM ] → panels, corners, ties, props, m², timber
- The LLM is the "eyes": it reads an image and turns it into numbers.
- The MCP is the "calculator": repeatable, no guessing. The same geometry always yields the same result - unlike LLM arithmetic.
Tool domains
| Domain | Status | Tools |
|---|---|---|
formwork |
ready | list_formwork_systems, formwork_catalog, count_formwork, concrete_pressure_check |
| (future) | planned | additional construction tools register the same way |
Formwork systems
| System | Max pressure | Catalog |
|---|---|---|
| BAUTEKK | 40 kN/m² | verified against DTR BauTekk 04/2022 + project PDF (R01) |
| BAUFRAME | 60 kN/m² | default values — to be verified |
| BAUSCHAL | 80 kN/m² | default values — to be verified |
BAUTEKK data taken from the manufacturer DTR: panel widths (10/20/25/30/45/50/60/75/90 cm, 70 cm as VT), three heights (90/120/150 cm), real article numbers for every size, outer/inner corners per height, and the connector rule (5/4/3 connectors per joint for 150/120/90 cm plates). Tie and prop counts remain design-dependent estimates.
BAUFRAME/BAUSCHAL catalogs contain sensible starting values. Confirm panel sizes and article numbers against the official BAUKRANE catalog before production use (
src/ged_invest_mcp/formwork/catalog.py).
Install
Requires Python ≥ 3.10.
python3.13 -m venv .venv
source .venv/bin/activate
pip install -e .
Run
The server supports two transports from a single entry point.
stdio (local — Claude Desktop / Cursor)
python -m ged_invest_mcp.server # stdio by default
ged-invest-mcp
In this mode you host no network server — the client (Claude/Cursor) launches the Python process itself and talks to it over stdin/stdout. When the client closes, the process ends. No ports, no hosting.
HTTP (remote — ChatGPT / connectors)
python -m ged_invest_mcp.server --http # http://0.0.0.0:8000/mcp
python -m ged_invest_mcp.server --http --port 9000
python -m ged_invest_mcp.server --transport sse # legacy SSE transport
ChatGPT cannot launch a local process, so it needs an MCP exposed over HTTP at a
URL. Expose a local server via a tunnel (e.g. ngrok http 8000) or deploy it,
then add https://.../mcp as a custom connector.
count_formwork — two ways to pass geometry
A) Segment list (you know wall lengths and corners):
{
"system": "BAUTEKK",
"segments": [
{"length": 465, "height": 150, "corner_at_end": "outer", "label": "A"},
{"length": 465, "height": 150, "corner_at_end": "outer", "label": "B"}
]
}
B) Polygon (outline coordinates — corners detected automatically):
{
"system": "BAUTEKK",
"polygon": [[0,0],[840,0],[840,700],[0,700]],
"height": 150
}
Client configuration
Claude Desktop / Cursor (mcp.json)
{
"mcpServers": {
"ged-invest": {
"command": "/ABSOLUTE/PATH/bud-ged-mcp/.venv/bin/python",
"args": ["-m", "ged_invest_mcp.server"]
}
}
}
Method (formwork, in short)
- Wall outline → segments between corners (outer / inner / none for collinear).
- Segment length → panel width selection (dynamic programming, min. timber).
- Height → panels are stacked to reach the wall height; the top course may
stand above the pour line (reported as
top_overshoot_cm, panels are not cut). - Panels = columns × courses × faces (default 2 = wall centerlines; pass
faces=1if the input already lists both faces separately). - Corners: one corner element per course.
- Hardware: full family from the catalog (ties, nuts, cones, plugs, spacer tubes, connectors, pins, wedges, corner tensioners, props with heads/feet). Coefficients are approximate and should be calibrated against the DTR.
- Horizontal leftover that does not fit panels → timber infill [lm].
Auditable output
count_formwork returns, besides the BOM:
input_echo— the geometry actually used (polygon-derived side lengths, corner types and winding), so you can verify the LLM read the drawing correctly;assumptions— faces, corner rule, spacings, panel heights, policies;- per-segment
panels_horizontal_per_face/courses_vertical_per_facelayout; reconciliation— geometry area vs panel area and waste %;units,catalog_version,warnings,disclaimer.
This is an engineering estimator — the result supports quoting/ordering and must be approved by the site manager (per the project notes).
concrete_pressure_check — fresh concrete pressure (DIN 18218)
Computes the maximum characteristic lateral fresh concrete pressure per
DIN 18218:2010-01 from the pouring (rise) rate v, consistency class
(F1–F6, SCC) and setting time tE, caps it by the hydrostatic pressure
γc·H, and compares it to the formwork system's limit:
characteristic_pressure_kn_m2,hydrostatic_pressure_kn_m2,design_pressure_kn_m2,governing;within_limitandmax_pouring_rate_for_limit_m_per_h(the fastest safe rate);- pass
system(e.g.BAUTEKK) to auto-fill the limit (40 kN/m²) and the DTR process cap (2 m/h).
Simplified 15 °C reference model (no temperature/admixture corrections). The real pressure and safe pouring rate remain the site manager's responsibility.
Adding a new tool domain
- Create
src/ged_invest_mcp/<domain>/with aregister(mcp)function. - Import and call it in
server.py.
Tests
python -m pytest -q
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。