DotnetTemplateMCP

An MCP server that lets AI agents work with dotnet new templates — search, inspect, preview, and create projects through natural conversation instead of memorizing CLI flags.

Instead of this:

dotnet new list --language C#
dotnet new webapi --help
dotnet new webapi --auth Individual --use-controllers --name MyApi --output ./MyApi

Your AI agent just says: "I need a web API with authentication and controllers" — and the MCP server figures out the rest.

Template Validation for Authors

Building a custom dotnet new template? template_validate catches mistakes before you publish — no more guessing if your template.json is correct:

Agent calls: template_validate("./my-template")

← Returns:
{
  "valid": false,
  "summary": "2 error(s), 1 warning(s), 3 suggestion(s)",
  "errors": [
    "Missing required field 'shortName'.",
    "Parameter 'Framework': default value 'net7.0' is not in the choices list."
  ],
  "warnings": [
    "Missing 'sourceName'. Without it, the generated project name won't be customizable via --name."
  ],
  "suggestions": [
    "Consider adding a 'description' field to help users understand what this template creates.",
    "Consider adding 'language' tag (e.g., 'C#') for better discoverability.",
    "Consider adding 'type' tag (e.g., 'project', 'item') for filtering."
  ]
}

What it catches: missing required fields, invalid identity format, short name conflicts with CLI commands, parameter issues (missing defaults, empty choices, prefix collisions, type mismatches), broken computed symbols, constraint misconfiguration, and missing tags.

No existing tooling does this — most template authors discover issues only after dotnet new install fails or produces wrong output.

Tools

Tool	What it does
template_search	Search locally and on NuGet.org — one call, ranked results
template_list	List what's installed, filter by language/type/classification
template_inspect	Parameters, constraints, post-actions — all in one shot
template_instantiate	Create a project. Not installed? Auto-resolves from NuGet. Elicits missing params interactively
template_dry_run	Preview files without touching disk
template_install	Install a package (idempotent — skips if already there)
template_uninstall	Remove a template package
templates_installed	Inventory of everything installed
template_from_intent	"web API with auth" → webapi + auth=Individual — no LLM needed
template_create_from_existing	Analyze a .csproj → generate a reusable template matching repo conventions
template_compose	Execute a sequence of templates (project + items) in one workflow
template_suggest_parameters	Suggest parameter values with rationale based on cross-parameter relationships
template_validate	Validate a local template directory for authoring issues before publishing
template_compare	Compare 2+ templates side by side — parameters, features, frameworks
solution_analyze	Analyze a solution/workspace — project structure, frameworks, CPM status

📖 Full tool reference →

Hosted deployment

A hosted deployment is available on Fronteir AI.

Quick Start

Global tool (.NET 8+)

dotnet tool install --global DotnetTemplateMCP --version 1.3.0

Zero-install with dnx (.NET 10+)

dnx -y DotnetTemplateMCP --version 1.3.0

VS Code / GitHub Copilot

Add to mcp.json:

{
  "servers": {
    "dotnet-templates": {
      "type": "stdio",
      "command": "dnx",
      "args": ["-y", "DotnetTemplateMCP", "--version", "1.3.0"]
    }
  }
}

📖 Claude Desktop, Cursor, and more →

Transport Modes

Stdio (default)

Standard I/O transport for local CLI and tool usage:

template-engine-mcp                     # stdio is the default
template-engine-mcp --transport stdio   # explicit

HTTP (remote / cloud / team-shared)

Streamable HTTP transport for remote, multi-tenant, or CI/CD deployment:

template-engine-mcp --transport http
# or via environment variable:
MCP_TEMPLATE_TRANSPORT=http template-engine-mcp

The HTTP server exposes:

• /mcp — MCP streamable HTTP endpoint
• /health — Health check endpoint

Configure the listen URL:

MCP_TEMPLATE_HTTP_URL=http://0.0.0.0:8080 template-engine-mcp --transport http

Connect your MCP client:

{
  "servers": {
    "dotnet-templates": {
      "type": "http",
      "url": "http://localhost:5005/mcp"
    }
  }
}

Interactive Elicitation

When a template has required parameters that weren't provided, the server asks the user interactively via MCP elicitation — instead of failing. Template parameter types are mapped to form fields:

Template Parameter	Elicitation Field
string	Text input
bool / boolean	Checkbox
int / number	Number input
Choice parameter	Single-select dropdown

Disable with MCP_TEMPLATE_ELICITATION=false.

How it works

You: "I need a web API with authentication, controllers, and Docker support"

→ template_from_intent extracts keywords: web api, authentication, controllers, docker
→ Matches: webapi (confidence: 0.85)
→ Resolves: auth=Individual, UseControllers=true, EnableDocker=true
→ template_instantiate creates the project

The server also does smart defaults (AOT → latest framework, auth → HTTPS stays on), parameter validation before writing files, constraint checking (OS, SDK, workload), interactive elicitation of missing required parameters, and auto-resolves templates from NuGet if they're not installed.

CPM & Latest Package Versions

When creating a project inside a solution that uses Central Package Management, the server automatically:

1. Detects Directory.Packages.props by walking up the directory tree
2. Strips Version attributes from generated .csproj PackageReferences
3. Adds missing <PackageVersion> entries to Directory.Packages.props
4. Resolves latest stable NuGet versions — no more stale hardcoded versions from templates

Before (what dotnet new generates):
  <PackageReference Include="Serilog" Version="3.1.0" />    ← stale, breaks CPM

After (what template_instantiate produces):
  .csproj:                    <PackageReference Include="Serilog" />
  Directory.Packages.props:   <PackageVersion Include="Serilog" Version="4.2.0" />

Works for standalone projects too — versions are updated directly in the .csproj.

Multi-Template Composition

Chain multiple templates in one call with template_compose:

[
  {"templateName": "webapi", "name": "MyApi", "parametersJson": "{\"auth\": \"Individual\"}"},
  {"templateName": "gitignore", "target": "."}
]

📖 Architecture & smart behaviors →

Tool Profiles (Lite vs Full)

By default, all 14 tools are available. If your agent works better with fewer tools, set the MCP_TEMPLATE_TOOL_PROFILE environment variable:

Profile	Tools	When to use
full (default)	All 14 tools	Full control — advanced workflows, composition, custom templates
lite	5 core tools	Simpler agents that just need to find and create projects

Lite profile tools: template_from_intent, template_instantiate, template_inspect, template_search, template_dry_run

{
  "servers": {
    "dotnet-template-mcp": {
      "command": "dotnet-template-mcp",
      "env": {
        "MCP_TEMPLATE_TOOL_PROFILE": "lite"
      }
    }
  }
}

Non-lite tools will return a helpful message explaining they're disabled and how to enable them.

Documentation

Doc	What's in it
Configuration	VS Code, Claude Desktop, Cursor setup + troubleshooting
Tool Reference	Every tool's parameters, types, and examples
Architecture	Template cache, smart behaviors, telemetry, project structure
MCP vs Skills	Why MCP over Copilot Skills — benefits and downsides
Plain LLM vs MCP	Side-by-side: what a plain LLM does vs. the MCP tool (4 scenarios)
Skills Equivalent	What it'd take to cover this with Skills instead

Building & Testing

dotnet build
dotnet test    # 185+ tests — unit, integration, and E2E

CI runs on push/PR via GitHub Actions (Ubuntu + Windows).

Contributing

Contributions are welcome! Please open an issue to discuss proposed changes before submitting a PR.

# Setup
dotnet restore
dotnet build

# Run tests
dotnet test

# Pack locally
dotnet pack src/Microsoft.TemplateEngine.MCP -o nupkg/

Changelog

See CHANGELOG.md for release history.

License

MIT