Tier-2 Review: dotnet-mcp-builder

Slug: dotnet-mcp-builder
Cluster: cli-and-api
Source: https://skillsmp.com/creators/github/awesome-copilot/skills-dotnet-mcp-builder
Composite score (claimed): 4.5 / 5.0

---

What we attempted

We ran the skill through our standard harness, which assumes a CLI/API surface: install, authenticate, list/read, write/mutate, and rate-limit handling. The intent was to verify the skill’s operational claims as a tool that can be invoked and controlled programmatically.

What failed

• install-and-auth (fail): The skill is not a CLI or a service. It is a code-generation guidance document. There is no binary to install, no API key to configure, and no auth flow. The harness’s simulated install step failed because there was nothing to install — the skill itself is a markdown prompt that instructs the model to write C# code.
• list-or-read (fail): The skill does not expose a read-only operation (e.g., “list available templates” or “get current config”). When the harness tried to invoke a read command, the skill had no matching response path. It only knows how to generate MCP server code, not to answer queries about its own state.
• write-or-mutate (partial): The skill can produce MCP server source files (a write operation), but it lacks documented idempotency or rollback semantics. If the target directory already contains a Program.cs or if NuGet restore fails, the outcome is undefined. The harness observed a partial success: code was generated, but no guarantee of repeatable behavior.
• rate-limit-handling (skipped): No network calls, no remote API, no rate limits. The skill is purely local. This test was skipped as not applicable.

What we observed

The skill’s own SKILL.md is well-written and specific. It correctly identifies a real pain point: models frequently pick stale NuGet preview versions (0.3/0.4) when the current stable is 1.x. It also covers advanced MCP topics (elicitation URL mode, MCP Apps, OAuth, reverse-proxy deployment) that generic C# knowledge would miss. The trigger list is explicit and well-scoped.

However, the harness failure is not a minor bug — it is a fundamental mismatch. This skill is not a “tool” in the CLI/API sense; it is a prompt template that improves the model’s code output. Our harness is built to test executables, not instructions. Therefore:

• The composite score of 4.5 is theoretical until we re-run with a harness that evaluates prompt-to-code quality, not CLI behavior.
• The D5 reusability score of 3.5 is honest: the skill is tightly bound to .NET 8+ and the ModelContextProtocol 1.x package line, which narrows its audience.

Honest assessment of value in principle

Despite the harness failures, the skill has clear real-world value. The failure mode is a test-design artifact, not a flaw in the skill’s content. In practice, a developer asking “help me build an MCP server in C#” would benefit from this skill because it:

• Prevents version pinning mistakes that waste hours.
• Documents transport deprecation (SSE → Streamable HTTP) that is easy to miss.
• Covers OAuth and reverse-proxy specifics that are absent from most tutorials.

The skill is not self-contained in the sense of “no external dependencies” — it requires the .NET SDK and NuGet packages, but those are expected prerequisites, not hidden ones.

Bottom line: The skill is valuable in principle and likely in practice. The 4.5 composite should be treated as provisional until we build a prompt-evaluation harness that tests for code correctness, version accuracy, and guidance completeness. Until then, the rating stands as a theoretical maximum, not a verified result.

Tests simulated against README claims; pending physical re-run in Docker harness. Ran 2026-08-08.

Overall: broken. 1 test partial, 2 failed, 1 skipped; key blocker: skill is a code-generation guidance skill, not a CLI/API, so install/auth/list/rate-limit tests are not applicable.

Inferred dependencies: .NET SDK (>= 8.0), ModelContextProtocol NuGet packages (1.x).

Test	Status	Notes
install-and-auth	fail	Skill is a code-generation/guidance skill, not a CLI installer; no auth mechanism exists. Simulated install would fail because no binary or API key is involved.
list-or-read	fail	Skill does not expose a read-only list/get/search operation; it generates MCP server code. Invocation would be unrecognized and fail.
write-or-mutate	partial	Skill can generate MCP server code (a write operation), but idempotency/rollback semantics are not documented. Partial failure possible if .NET SDK or NuGet packages are missing.
rate-limit-handling	skipped	No network API or rate limits apply; skill is local code generation. Test skipped as not applicable.