PyNet Bridge
🐍 PyNet 平台桥接器 (MCP)
PyNet 平台桥接器 (MCP) 是一个执行层,允许 AI 模型实时控制 Autodesk 工具。
它连接了自然语言 → Python → Autodesk 桌面工具(Navisworks、Revit、AutoCAD),使 AI 能够自主生成、执行和优化 BIM 工作流程。
支持的集成包括 Navisworks Manage、Revit 和 AutoCAD。
该桥接器充当 AI 逻辑与 Autodesk 桌面 API 之间的连接纽带,允许使用自然语言进行动态 UI 创建、脚本执行和 BIM 流程自动化。
🔄 工作原理
用户用自然语言描述任务。
AI 生成 Python 脚本。
PyNet Bridge 验证并发送脚本。
PyNet 插件在 Autodesk 内部执行脚本。
结果返回给 AI。
这就是将 AI 从聊天机器人转变为 BIM 执行引擎的关键。
Related MCP server: A2A MCP Server
🚀 PyNet Bridge 的强大之处
AI → 行动: 将 AI 生成的代码转化为 Navisworks/Revit 内部的实际操作
实时执行: 无需离开 BIM 环境即可立即运行脚本
动态 UI 创建: 让 AI 随时创建工具、按钮和工作流程
可靠的通信: 快速且稳定的本地 IPC
模型感知自动化: 直接在实时 BIM 模型上运行
🛠️ 安装
✅ 选项 A — 自动安装程序(推荐)
打开 PowerShell 并运行:
irm https://raw.githubusercontent.com/rafa2403nunez-droid/PyNetBridge/main/install.ps1 | iex这将自动执行以下操作:
检查是否安装了 Python 3.10+
从 PyPI 安装
pynet-mcp-bridge(通过uv或pip)自动检测并配置所有已安装的 AI 客户端:
Claude Desktop(标准版和 Microsoft Store 版本)
Claude Code(VS Code 扩展 / CLI)
Cline(VS Code 扩展)
Roo Code(VS Code 扩展)
pynet-mcp-bridge 软件包包含:
软件包 | 用途 |
pynet-mcp-bridge | 通过 PyNET 将 AI 模型与 Autodesk Navisworks 连接的 MCP 服务器 |
mcp[cli] | 模型上下文协议 (MCP) SDK 和 CLI 工具 |
fastmcp | 高级 MCP 服务器框架 |
psutil | 系统进程检测(查找正在运行的 Autodesk 实例) |
安装后请重启您的 AI 客户端以应用更改。
📦 Python 库入门包(可选)
安装推荐的 Python 库,以便使用 PyNET 进行 Navisworks/Revit 脚本编写:
irm https://raw.githubusercontent.com/rafa2403nunez-droid/PyNetBridge/main/install-libraries.ps1 | iex这将安装:
库 | 用途 |
pandas | 数据分析与处理 |
plotly | 交互式图表与可视化 |
matplotlib | 静态绘图与图表 |
dash | 基于 Python 的 Web 仪表板 |
这些是 允许的 Python 导入 中列出的第三方库。标准库模块(
json、sys、re等)已包含在 Python 中。
先决条件
PyNet Platform 插件已安装在 Navisworks/Revit 中。
Python 3.10 或更高版本 → python.org
⚠️ 暂不支持 Python 3.14。
pythonnet运行时目前支持 Python 3.7 到 3.13。如果您遇到提及不支持的 ABI 版本的System.NotSupportedException,请安装 Python 3.12 或 3.13 并配置 PyNet 使用它。uv → docs.astral.sh/uv
Git → git-scm.com — VS Code 扩展(Claude Code、Cline、Roo Code)正常运行所必需。
对于 Cline / Roo Code:VS Code → code.visualstudio.com
🔧 选项 B — 手动安装
1. 安装软件包:
uv tool install pynet-mcp-bridge或使用 pip:
pip install pynet-mcp-bridge2. 配置 Claude Desktop:
将以下内容添加到您的 claude_desktop_config.json 中:
标准版:
%APPDATA%\Claude\claude_desktop_config.jsonMicrosoft Store 版:
%LOCALAPPDATA%\Packages\Claude_*\LocalCache\Roaming\Claude\claude_desktop_config.json
{
"mcpServers": {
"pynet-bridge": {
"command": "pynet-bridge",
"args": []
}
}
}3. 配置 Claude Code (VS Code 扩展):
添加到 %USERPROFILE%\.claude.json:
{
"mcpServers": {
"pynet-bridge": {
"type": "stdio",
"command": "pynet-bridge",
"args": []
}
}
}4. 配置 Cline:
添加到 %APPDATA%\Code\User\globalStorage\saoudrizwan.claude-dev\settings\cline_mcp_settings.json:
{
"mcpServers": {
"pynet-bridge": {
"type": "stdio",
"command": "pynet-bridge",
"args": []
}
}
}5. 配置 Roo Code:
添加到 %APPDATA%\Code\User\globalStorage\rooveterinaryinc.roo-cline\settings\mcp_settings.json:
{
"mcpServers": {
"pynet-bridge": {
"type": "stdio",
"command": "pynet-bridge",
"args": []
}
}
}🛠️ 可用的 MCP 工具
这些工具允许 AI 完全控制 PyNet 环境,从 UI 创建到脚本执行和系统监控。 连接后,AI 将能够访问全套 PyNet 工具:
🧠 暴露给 AI 的核心能力
🔍 系统与连接
list_active_instances:扫描系统以查找具有活动 PyNet 连接的正在运行的 Autodesk 进程。
check_plugin_status:握手 ping 以验证插件监听器是否响应。
🏗️ 模块 (选项卡) 管理
get_pynet_ui_layout:获取完整的 UI 结构(ButtonsModules 和 ScriptButtons)。
create_pynet_module:在功能区中创建一个新的自定义选项卡 (ButtonsModule)。
delete_pynet_module:永久删除模块及其所有内容。
🔘 按钮管理
get_buttons_data:列出特定模块 ID 的所有脚本按钮。
deploy_script_button:将新的 ScriptButton 安装到特定模块中(名称、脚本、图标、工具提示)。
update_script_button:更新现有 ScriptButton 的元数据或将其移动到另一个模块。
delete_script_button:按 ID 从模块中永久删除 ScriptButton。
💻 执行与控制台控制
send_command:在 PyNet 引擎中直接执行脚本(目标 PID、脚本名称、内容)。
get_output_window_status:检查输出窗口当前是否可用/可见。
configure_output_window:切换 PyNet 日志/输出窗口的可见性。
🛡️ 安全的 AI 执行
PyNet Bridge 包含一个内置的验证层,确保所有 AI 生成的脚本在执行前都是安全且受控的。
✔ 防止不安全的操作 ✔ 阻止未经授权的系统访问 ✔ 保证与 BIM 模型的受控交互
AI 依然强大,但处于安全边界内
从 v1.1.1 开始,MCP 服务器包含一个内置的静态分析器,在脚本到达 Autodesk 主机之前对其进行验证。所有脚本都在桥接器级别进行解析和检查 — 被拒绝的脚本永远不会离开 MCP 服务器。
允许的 CLR 程序集
仅允许通过 clr.AddReference 使用以下 .NET 引用:
通用:
System,System.Windows.Forms,System.Drawing,System.Collections.GenericNavisworks:
Autodesk.Navisworks.Api,.ComApi,.Interop.ComApi,.ClashRevit:
RevitAPI,RevitAPIUIAutoCAD / Civil 3D:
AcMgd,AcCoreMgd,AcDbMgd,AecBaseMgd,AecPropDataMgd,AeccDbMgdPyNet 插件:
Raen.Core.Pynet.*,Raen.{Product}.Pynet.*(任何版本 — 例如Raen.Core.Pynet.Resources,Raen.Navisworks.Pynet.2024,Raen.Civil3D.Pynet.2026)
允许的 Python 导入
clr, sys, json, re, time, datetime, pathlib, typing, threading, collections, xml, pandas, plotly, matplotlib, dash, webbrowser, psutil, functools
允许的 Python 子模块
某些模块仅在子模块级别允许,以防止访问危险的同级模块:
允许 | 阻止 | 原因 |
|
| 允许本地 HTTP 服务,阻止出站请求 |
阻止的 Python 导入
os, subprocess, shutil, socket, ctypes, pickle, importlib, urllib, signal, multiprocessing, tempfile, glob, inspect, code, codeop
阻止的调用
eval, exec, compile, __import__, getattr, setattr, delattr, globals, locals, vars, breakpoint, open
阻止的属性访问
__builtins__, __subclasses__, __globals__, __code__
任何违反这些规则的脚本都会立即被拒绝并显示描述性错误消息,且永远不会发送到插件。
📂 项目结构
pynet_mcp/:核心 MCP 服务器逻辑 (FastMCP)。
pyproject.toml:软件包配置和依赖管理。
📥 入门指南
在几分钟内开始构建自主 BIM 工作流程。
安装桥接器,连接您的 AI 客户端,并将自然语言转化为模型中的实际操作。
❓ 常见问题解答
对安装、配置或使用有疑问?查看完整的常见问题解答页面:
🔗 此 MCP 如何融入生态系统
此 MCP 是一个模块化系统的一部分,旨在实现跨 Autodesk 工具的 AI 驱动 BIM 自动化。
此存储库旨在与以下内容协同工作:
PyNet Platform → 通过 Python.NET 在 Navisworks、Revit 和 Civil 3D 中执行脚本。
PyNet Library → 通过 Python 脚本库为 AI 提供上下文。
这些组件共同实现了:
自然语言 → AI → Python 脚本 → PyNet → Autodesk → BIM 操作
组件 | 存储库 | 用途 |
PyNet Platform | Navisworks、Revit 和 Civil 3D 插件 — 托管 Python.NET 引擎 | |
PyNet Bridge (MCP) | 本存储库 | MCP 服务器 - 将 AI 模型连接到 PyNET,并包含安全的脚本验证 |
PyNet Library | 脚本参考库和 AI 上下文 |
📄 许可证
本项目采用 MIT 许可证授权。
Available Tools
24 toolscheck_plugin_statusBRead-only
Handshake ping to verify the plugin listener is responsive.
Inputs: pid (required). Returns: result
configure_output_windowCDestructive
Toggles the visibility of the PyNet log/output window.
Inputs: pid (required), is_available (required). Returns: result
create_pynet_moduleCDestructive
Creates a new custom Tab (ButtonsModule) in the Ribbon.
Inputs: pid (required), name (required). Returns: result
delete_pynet_moduleCDestructive
Permanently deletes a module and all its contents.
Inputs: pid (required), module_id (required). Returns: result
delete_script_buttonBDestructive
Permanently removes a ScriptButton from a module by Id.
Inputs: pid (required), button_id (required), module_id (required). Returns: result
deploy_script_buttonBDestructive
Installs a new ScriptButton into a specific module (Name, Script, Icon, Tooltip).
Inputs: pid (required), name (required), module_id (required), script_Path (required), +2 more. Returns: result
get_buttons_dataARead-only
Lists all script buttons for a specific module ID.
Inputs: pid (required), module_id (required). Returns: result
get_output_window_statusBRead-only
Checks if the output window is currently available/visible.
Inputs: pid (required). Returns: result
get_pynet_ui_layoutCRead-only
Fetches the full UI structure (ButtonsModules and ScriptButtons).
Inputs: pid (required). Returns: result
list_active_instancesARead-only
Scans the system for running host processes (Autodesk, Rhino, AVEVA UE, PyNET for Tekla/Presto) with an active PyNet IPC pipe.
Inputs: no named parameters. Returns: result
send_commandBDestructive
Direct script execution in the PyNet engine (Target PID, Script Name, Content).
Inputs: pid (required), content (required), timeout (required), script_name (required). Returns: result
send_command_by_pathADestructive
Executes a script file directly by path in the PyNet engine, without sending content inline.
Inputs: pid (required), timeout (required), file_path (required), script_name (required). Returns: result
update_script_buttonBDestructive
Updates metadata for an existing ScriptButton or moves it to another module.
Inputs: pid (required), name (required), tooltip (required), button_id (required), +4 more. Returns: result
viewer_clearARead-only
Clears highlights and ghosting in the open viewer.
Inputs: no named parameters. Returns: result
viewer_fitARead-only
Fits the camera to all models in the open viewer.
Inputs: no named parameters. Returns: result
viewer_get_propertiesARead-only
Reads element properties (name, model, psets) from the loaded package's properties.json.
Pass one or more pnt_ids to get their full properties (including psets) directly.
Called with no pnt_ids, returns a paginated lightweight index (pnt_id, name, model) instead
of every element — narrow it with model (exact model/discipline name, e.g. "Snowdon Towers
Sample HVAC") and/or search (case-insensitive substring of the element name), and page
through with limit/offset (default 200 per page, capped at 500). On a federated model
the unfiltered index can have tens of thousands of entries — always pass model or search
when you can, and use offset to page through the rest.
The data source is the .pnt's properties.json, NOT the viewer or IFC.
Inputs: limit, model, offset, search, +1 more. Returns: result
viewer_get_stateARead-only
Reads the viewer's last reported state: loaded models and the pnt_id of whatever the user currently has selected (click in the 3D view or the tree panel — null if nothing is selected). Use the returned pnt_id with viewer_get_properties for its details, or feed it into a clashes.json lookup to find what it clashes with.
Inputs: no named parameters. Returns: result
viewer_highlight_clashADestructive
Highlights a clash (element(s) A red, element(s) B green) and isolates the rest away.
Usually one pnt_id per side (a single clash pair), but either side accepts a list — e.g. one element (A) against every counterpart it clashes with (B), all shown at once.
Inputs: pnt_id_a, pnt_id_b. Returns: result
viewer_isolateADestructive
Isolates (hides everything except) the given pnt_ids in the open viewer.
Inputs: pnt_ids. Returns: result
viewer_isolate_modelsADestructive
Shows only the given model(s)/discipline(s) (by exact name, e.g. "Snowdon Towers Sample HVAC"), hiding every other loaded model. Cheaper than viewer_isolate for "just show me this discipline" requests — no need to resolve every element's pnt_id first. Get the exact model names from viewer_get_state's "models" list. Undo with viewer_clear.
Inputs: models. Returns: result
viewer_list_clashesARead-only
Reads the loaded package's clashes.json — the data source, NOT the viewer/IFC.
Returns each clash with the pnt_id identifiers needed to highlight it.
Inputs: no named parameters. Returns: result
viewer_load_packageBDestructive
Loads a .pnt into the open viewer and returns a summary read from clashes.json.
Inputs: pnt_path (required). Returns: result
viewer_selectADestructive
Highlights a set of elements in the open viewer by pnt_id, in neutral colours (yellow / blue) with nothing else hidden — a plain "look at these" pointer.
Pass a list of pnt_ids in pnt_ids (group A). Optionally pass a second list in
pnt_ids_b to highlight a second group in a distinct colour. Replaces any previous
selection. Use viewer_list_clashes / viewer_get_properties to obtain pnt_ids.
For a clash pair (red/green, with the rest of the model isolated away), use
viewer_highlight_clash instead.
Inputs: pnt_ids, pnt_ids_b. Returns: result
viewer_statusARead-only
Reports whether a PyNet BIM Viewer is open in VS Code (port, package, data dir).
Inputs: no named parameters. Returns: result
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 24 tools
The viewer tools have the most potential for confusion (select vs highlight_clash vs isolate vs isolate_models vs clear), but the descriptions explicitly cross-reference each other and state when to prefer one over another, which resolves most of it. A few boundaries remain soft — check_plugin_status vs list_active_instances vs viewer_status all report 'is something alive', and configure_output_window vs get_output_window_status are near-mirror pairs — but they are separable.
Nearly everything follows verb_noun (get_, create_, delete_, update_, deploy_, list_, send_, configure_, check_), with a clean viewer_ namespace for the 3D side. Minor deviations: the ribbon/button noun is inconsistent ('buttons' in get_buttons_data vs 'script_button' in deploy/update/delete_script_button, and get_pynet_ui_layout is a noun-first outlier), and some tools carry a domain prefix while others do not.
24 tools is on the heavy side, and the surface spans four separate concerns (viewer control, ribbon/module management, IPC process discovery, script execution). The count is defensible given the breadth of the PyNet bridge, but a few low-value entries (viewer_fit, get_output_window_status) could be folded into neighbours.
Script buttons have full CRUD (deploy/update/delete/get_buttons_data), the viewer has load/state/properties/select/isolate/highlight/clear/fit, and execution has both inline and by-path variants. The main gap is module lifecycle: create_pynet_module and delete_pynet_module exist but there is no update/rename/reorder module tool, and there is no way to unload a package or detach from a running instance.
Maintenance
Related MCP Connectors
A2a Governance Bridge MCP Server by MEOK AI Labs
MCP federation gateway: route, govern and observe agent tools across clouds and MCP servers.
MCP Server for an Agent Task Marketplace
Telegram bridge for your MCP-compatible agent. Bidirectional, no LLM in our stack.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceControl AI agents creation and modification via MCP.-
- AlicenseAqualityFmaintenanceA bridge server that enables MCP-compatible AI assistants like Claude to seamlessly discover, communicate with, and manage A2A protocol agents.7113 PyPI146Apache 2.0
- AlicenseBqualityFmaintenanceA bridge server that connects Agent Communication Protocol (ACP) agents with Model Context Protocol (MCP) clients, enabling seamless integration between ACP-based AI agents and MCP-compatible tools like Claude Desktop.1665 PyPI24MIT
- AlicenseNot gradedqualityDmaintenanceMCP server for cross-platform agent onboarding. Registers external agents, translates intents from LangChain, CrewAI, AutoGen, and A2A formats, and proxies cross-ecosystem transactions.MIT