MCP server for evaluating and executing C# scripts using Roslyn, with support for NuGet package references
Single tool (EvalCSharp) with a clear verb-based name and comprehensive description. Parameters have types and descriptions, and basic schema structure is present. However, the output schema is completely undocumented (returns raw string), error handling lacks actionable guidance for recovery, and security constraints around code execution are mentioned but not enforced at the tool interface level. The tool is high-risk (IRREVERSIBLE) but lacks confirmation/dry-run patterns. Descriptions are adequate (194 chars for tool, ~50-70 chars per param) but lack examples of expected output format or error conditions.
Evaluates and executes C# script code and returns the output. Can either execute code directly or from a file. Supports NuGet package references using #r "nuget: PackageName, Version" directives.
Output schema completely undocumented. Tool returns string with no declared structure for success vs. error cases. LLM cannot reliably parse whether execution succeeded, extract error details, or plan downstream calls.
No confirmation or dry-run pattern for irreversible operation. EvalCSharp executes arbitrary C# code with side effects (file I/O, network calls, assembly loading). LLMs can trigger unintended code execution without review.
Error handling returns free-form strings with no structured categorization. LLM cannot distinguish retryable errors (timeout) from fatal ones (syntax error) or user-fixable ones (missing file). Error messages lack actionable recovery hints.
No pagination or result limit documented. If script output is very large (e.g., printing millions of lines), it will blow the context window. Tool description should state a max output size and truncation behavior.
Inferred effective spec: <=2025-11-25.
| Scored | Grade | Overall | Spec posture | Rubric |
|---|---|---|---|---|
| 2026-09-22 | D | 50 | <=2025-11-25 | v2 |
| 2026-03-09 | F | 15 | - | v1 |
Security mechanism (CSX_ALLOWED_PATH environment variable) is not exposed as tool-level configuration. Path validation is implementation detail hidden from LLM. Tool description does not clarify what directories are accessible or how to request access.
Parameter 'timeoutSeconds' has a default (30s) but no documented min/max range. LLM could pass 0, negative values, or 999999 without constraint. Baseline from rubric: numeric params should specify range (e.g., 1 - 300).
Mutual exclusion of 'csxFile' and 'csx' is documented in error message only, not in parameter descriptions. LLM may attempt to pass both parameters, causing a rejection. Should explicitly state: 'Exactly one of csxFile or csx is required.'