Civil 3D MCP Server
Enables AI assistants to write and execute C# code directly inside Autodesk Civil 3D, using read-only and write tools plus reusable skill templates for drawing and civil object tasks.
README
Civil 3D MCP Server — Dynamic Roslyn Fork
An MCP server that enables AI assistants to write and execute C# code directly inside Autodesk Civil 3D. Instead of a large set of fixed tools, the AI generates task-specific code that runs with Civil 3D API access.
Project Scope and Lineage
This fork keeps the dynamic Roslyn/C# execution model and a deliberately small public surface of three MCP tools. Its current compatibility baseline is Autodesk Civil 3D 2025, with local work focused on reliability, safety, measurable efficiency, and reusable Civil 3D skills. Other Civil 3D versions can be added later through separately verified compatibility work.
The project is derived from barbosaihan/civil3d-mcp. SantosSjba/mcp-to-c3d was evaluated for selected, test-backed ideas, while Sacred-G/Civil3D-mcp was used only as an architectural reference. See PROVENANCE.md for the detailed attribution and licensing boundaries.
This independent project is not affiliated with or endorsed by Autodesk. Autodesk assemblies and other proprietary Civil 3D files are not included.
Architecture
┌─────────────────┐ stdio ┌──────────────────┐ TCP/JSON-RPC ┌──────────────────┐
│ AI Assistant │ ◄────────────► │ MCP Server (TS) │ ◄──────────────────► │ Civil 3D Plugin │
│ (Claude, Cline) │ │ 3 meta-tools │ port 8080 │ Roslyn Engine │
└─────────────────┘ └──────────────────┘ └──────────────────┘
│ │
Skills Library C# Code Execution
(.skill.md files) (full Civil 3D API)
3 Meta-Tools
| Tool | Purpose | Safety |
|---|---|---|
civil3d_execute |
Execute C# code with write access; optional save after commit | ⚠️ Modifies drawing |
civil3d_query |
Execute C# code read-only (no commit) | ✅ No side effects |
civil3d_skills |
Browse/search/read code skill templates; api_lookup searches already-loaded public Civil 3D API metadata |
✅ Metadata only |
How It Works
- AI reads a skill → Gets a documented C# code template
- AI adapts the code → Fills in parameters, combines patterns
- AI sends code → Via
civil3d_executeorcivil3d_query - Roslyn compiles + runs → Inside Civil 3D with full API access
- Results return as JSON → Back to the AI
Example Interaction
User: "What surfaces are in my drawing?"
AI: Uses civil3d_query with:
var surfaces = new List<object>();
foreach (ObjectId id in CivilDoc.GetSurfaceIds()) {
var s = Transaction.GetObject(id, OpenMode.ForRead) as TinSurface;
surfaces.Add(new { s.Name, s.Layer });
}
return surfaces;
Result: [{ "Name": "EG", "Layer": "C-TOPO-EG" }, ...]
Skills Library
civil3d_skills also supports action: "api_lookup" for a bounded, read-only search of public type and member names/signatures from already-loaded allowlisted Civil 3D host assemblies. It does not load assemblies, run C# code, or access the active drawing. Supply a query and optionally an assembly, namespace prefix, and result limit.
Skills are documented C# code templates in skills/:
skills/
├── surfaces/ # Surface operations
├── alignments/ # Alignment + station/offset
├── points/ # COGO points
├── geometry/ # Lines, polylines, text
├── drawing/ # Drawing info
└── workflows/ # Complex multi-object operations
Script Globals
Code executed via civil3d_execute or civil3d_query has access to:
| Global | Type | Description |
|---|---|---|
Document |
Document |
Active AutoCAD document |
CivilDoc |
CivilDocument |
Active Civil 3D document |
Database |
Database |
Document database |
Transaction |
Transaction |
Active transaction |
Editor |
Editor |
Document editor |
All Civil 3D namespaces are auto-imported.
Committing the civil3d_execute transaction changes the open drawing but does not by itself write the DWG file to disk. Set saveDrawing: true when the completed change should also be saved. The plugin saves only after the script transaction and document lock are closed; scripts must not call Database.SaveAs or queue QSAVE themselves. The save request uses a separate 10-minute default timeout and is never retried automatically.
Setup
1. Build MCP Server
npm install && npm run build
2. Build Plugin
# Copy DLLs from Civil 3D to C_References/ (see C_References/README.md)
cd plugin/Civil3dMcpPlugin
dotnet build
3. Load in Civil 3D
NETLOAD → select Civil3dMcpPlugin.dll
C3DMCPSTATUS → verify running
4. Configure AI
{
"mcpServers": {
"civil3d": {
"command": "node",
"args": ["/path/to/civil3d-mcp/build/index.js"]
}
}
}
Environment Variables
| Variable | Default | Description |
|---|---|---|
CIVIL3D_HOST |
localhost |
Plugin host |
CIVIL3D_PORT |
8080 |
Plugin port |
CIVIL3D_COMMAND_TIMEOUT |
120000 |
Execution timeout (ms) |
CIVIL3D_SAVE_TIMEOUT |
600000 |
Timeout for execute requests with saveDrawing: true (ms) |
LOG_LEVEL |
info |
Log level |
Benchmarking
The phase 2A host-independent recorder, the phase 2A.1 opt-in internal live trace contract, and the phase 2A.2 read-only live runner are documented in benchmark/README.md. None adds an MCP tool, queue, or retry; the 2A.2 runner can invoke only its fixed read-only query when explicitly started.
Structured errors (phase 2B.1)
civil3d_query and civil3d_execute keep their existing text error content and isError: true, while also returning structuredContent with schema civil3d-mcp-error/v1. The stable error fields are code, category, message, source, outcome, and retryable. A command timeout or a connection loss after sending has outcome: "unknown" and retryable: false; the server never retries it automatically. Successful responses and the three-tool public surface are unchanged.
Private TCP framing (phase 2C.1)
Each localhost TCP connection carries one UTF-8 JSON-RPC request and one response. Each JSON body is followed by LF and is limited to 8 MiB, measured as UTF-8 bytes without the LF. The Node client still accepts the previous plugin's unframed response when that complete JSON body is followed by an orderly connection close. Oversized requests are rejected before they are written; oversized or malformed responses and interrupted connections produce non-retryable structured transport errors. If execution completed but the plugin could not return an oversized result, the reported outcome is unknown.
Operation audit logging and write idempotency (phase 2I.1 / 2I.2)
At the default info log level, each accepted civil3d_query and civil3d_execute operation emits one bounded stderr audit event. It contains a new opaque operation ID, tool name, SHA-256 and UTF-8 byte length of the C# source, success/error status, and elapsed milliseconds; errors add only stable code/category/source/outcome fields. The audit event never contains caller code, description, drawing identity, result, or error message.
civil3d_execute also accepts an optional opaque idempotencyKey (1–128 ASCII letters, digits, ., _, :, -). In one plugin session it binds the key to the UTF-8 C# SHA-256, normalized expectedDrawing identity, and saveDrawing choice. A duplicate is rejected as in-progress, conflicting, or already committed; committed entries retain no result and callers must reconcile with a read-only query. A save failure occurs after the in-memory write commit, so its key is kept as completed to prevent an accidental duplicate modification. The session keeps at most 256 completed keys, evicting the oldest deterministically. This adds neither persistence nor automatic retry or exactly-once semantics.
Security
The Roslyn sandbox blocks:
- Process execution (
Process.Start) - File deletion (
File.Delete) - Network requests (
HttpClient,Sockets) - Registry access
- Dynamic assembly loading
All Civil 3D API operations are allowed.
This regex sandbox is defense in depth, not a trust boundary. Both code tools receive mutable Civil 3D and AutoCAD API objects; civil3d_query skips the host's transaction commit but cannot guarantee that arbitrary dynamic C# is side-effect-free. Run only trusted, approval-gated code. Loopback TCP prevents remote network access but does not authenticate other local processes.
License
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。
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。
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。
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。
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。
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。
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。
mcp-server-qdrant
这个仓库展示了如何为向量搜索引擎 Qdrant 创建一个 MCP (Managed Control Plane) 服务器的示例。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。