SharpLens
io.github.pzalutski-pixel由 GitHub 账号 pzalutski-pixel 发布v1.5.3更新于 Oct 26, 2026
67 semantic C#/.NET analysis tools via Roslyn: navigation, refactoring, codegen, audit.
概览
本地 MCP 服务器,借助 Roslyn 为 AI 助手提供 C#/.NET 解决方案的语义分析:导航、重构、代码生成与代码审计。
- 可以做什么
- 导航与搜索符号:定义、引用、实现、调用方、调用图、类型层次、重载、特性与 XML 文档
- 分析代码:编译器与分析器诊断、变更影响、数据流与控制流、复杂度指标、死代码、循环依赖、公共 API 差异
- 重构与生成:重命名、修改签名、提取方法或接口、整理 using、格式化、应用代码操作、添加空值检查、相等性成员与测试桩
- 质量审计:上帝对象、未被测试的代码、异步问题、架构规则、命名违规、未使用的依赖
- 检查工作区:加载解决方案、同步已编辑文件、查看项目结构、依赖图、NuGet 引用与生成的代码
- 适用场景
- 适合助手在 C#/.NET 代码库上工作、需要纯文本搜索或 LSP 无法给出的编译器级准确答案时使用,例如影响分析、死代码、复杂度指标,或带预览的解决方案级安全重构。
- 运行要求
- 以 stdio 在本地运行,可通过 .NET 全局工具或 npx 安装。需要 .NET 8.0 SDK 或更高版本以及支持 MCP 的客户端。需将 DOTNET_SOLUTION_PATH 设为 .sln、.slnx 或 .slnf 文件,或先调用 load_solution。可选设置包括 SHARPLENS_LOG_LEVEL、SHARPLENS_TIMEOUT_SECONDS、SHARPLENS_MAX_DIAGNOSTICS、SHARPLENS_ABSOLUTE_PATHS 与 SHARPLENS_ENABLE_SEMANTIC_CACHE。未声明任何身份验证。
安装
在 SourceWeft 中
- 打开 控制台中的 SharpLens,将其添加到工作区。
- 为需要使用其工具的对话启用该服务。
Desktop only,通过 STDIO。 STDIO 服务会启动本地进程,因此需要 SourceWeft 桌面宿主。
其他 MCP 客户端
参照 仓库 中的启动说明。
README
SharpLensMcp
A Model Context Protocol (MCP) server providing 92 AI-optimized tools for .NET/C# semantic code analysis, navigation, refactoring, and code generation using Microsoft Roslyn.
Built for AI coding agents - provides compiler-accurate code understanding that AI cannot infer from reading source files alone.
Installation
Via NuGet (Recommended)
Then run with:
Via npm
Build from Source
Claude Code Setup
- Install the tool (pick one):
- Create
.mcp.jsonin your project root:
-
Restart Claude Code to load the MCP server
-
Verify by asking Claude to run a health check on the Roslyn server
Why Use This with Claude Code?
Claude Code has native LSP support for basic navigation (go-to-definition, find references). SharpLensMcp adds deep semantic analysis:
Configuration
The pre-1.6.0 ROSLYN_* spellings of these variables are still read as a fallback; the SHARPLENS_* spelling wins when both are set. Boolean values are case-insensitive. A number that does not parse or is below its minimum falls back to the default, and one above its maximum is capped; each case logs a warning to stderr.
Multi-targeted projects
A project built for several target frameworks is one project to every tool: each result is reported once, with a targetFrameworks list where it exists only under some frameworks. projectName accepts the plain name (MyLib) for all frameworks, or a framework-qualified name (MyLib(net8.0)) for one.
If DOTNET_SOLUTION_PATH is not set, you must call the load_solution tool before using other tools.
Migrating from 1.5.x tool names
Tool names no longer carry the roslyn: prefix — the colon violates the MCP tool-name pattern (^[a-zA-Z0-9_-]{1,64}$), which some clients enforce. Every tool keeps its name minus the prefix:
tools/list publishes only the new names. Calls using the old prefixed names are still accepted as aliases for one release and will be removed in the following one.
AI Agent Configuration Tips
AI models may have trained bias toward using their native tools (Grep, Read, LSP) instead of MCP server tools, even when SharpLensMcp provides better capabilities.
To ensure optimal tool usage:
-
Claude Code: Add to your project's
CLAUDE.md: -
Other MCP clients: Configure tool priority in your agent's system prompt
The semantic analysis from Roslyn is more accurate than text-based search, especially for overloaded methods, partial classes, and inheritance hierarchies.
Agent Responsibility: Document Synchronization
Important: SharpLensMcp maintains an in-memory representation of your solution for fast queries. When files are modified externally (via Edit/Write tools), the agent is responsible for synchronizing changes.
When to call sync_documents:
Usage:
A full sync also picks up new .cs files the project's compile items include. SharpLensMcp's apply tools write only the files an edit changes, keep each file's encoding and line endings, and refuse with STALE_DOCUMENT when a target file changed on disk since it was read — run sync_documents, then retry. They also refuse with UNSUPPORTED_ENCODING to write a file whose bytes its encoding cannot reproduce (a BOM-less file that is not valid UTF-8 on a system whose fallback code page is UTF-8, such as Linux); save it as UTF-8 first.
Why this design?
This mirrors how LSP (Language Server Protocol) works - the client (editor) notifies the server of changes. This approach:
- Eliminates race conditions (agent controls timing)
- Avoids file watcher complexity and platform quirks
- Is faster than full solution reload
- Gives agents explicit control over workspace state
If you don't sync: Queries may return stale data (old method signatures, missing new files, etc.)
Features
- 92 Semantic Analysis Tools - Navigation, refactoring, code generation, diagnostics, discovery, audit/quality
- AI-Optimized Descriptions - Clear USAGE/OUTPUT/WORKFLOW patterns
- Structured Responses - Consistent
success/error/dataformat withsuggestedNextTools - Zero-Based Coordinates - Clear warnings to prevent off-by-one errors
- Preview Mode - Safe refactoring with preview before apply
- Batch Operations - Multiple lookups in one call to reduce context usage
Tool Categories
Navigation & Discovery (24 tools)
Analysis (17 tools)
Refactoring (16 tools)
Code Generation (3 tools)
Compound Tools (7 tools)
Audit & Quality (10 tools)
Discovery (3 tools)
Infrastructure (12 tools)
Other MCP Clients
For MCP clients other than Claude Code, add to your configuration:
Usage
- Load a solution: Call
load_solutionwith path to.slnor.slnxfile (or setDOTNET_SOLUTION_PATH) - Analyze code: Use any of the 92 tools for navigation, analysis, refactoring, audit
- Refactor safely: Preview changes before applying with
preview: true
Architecture
Requirements
- .NET 8.0 SDK or later — works with .NET 8, 9, 10, and future versions. Analyzes any .NET 8+ project/solution.
- MCP-compatible AI agent
FAQ
Why does the tool target net8.0 — can it analyze my .NET 9 / .NET 10 project?
Yes. net8.0 is the tool's own runtime floor — the Roslyn 5.x packages it builds on require it — not a ceiling on what it can analyze. RollForward lets the installed tool run on newer .NET runtimes, and MSBuildWorkspace loads each project's real target framework from its csproj, so one install analyzes solutions targeting .NET 8, 9, 10, and beyond.
Development
Adding New Tools
- Add the method to the matching
src/RoslynService.*.cspartial (Navigation, Analysis, Refactoring, CallAnalysis, …) and return through the shared response envelope:
-
Register one
ToolDefinitioninsrc/ToolRegistry.cs— its name, description, input schema, theReadOnlyflag (mutating tools passReadOnly: falseand receive adestructiveHintannotation), and a handler that binds arguments throughJsonRpcParametersand calls your method. The registry drives bothtools/listand dispatch; there is no separate switch to edit. Two tests keep it honest:ToolsListGoldenTestslocks the published schema byte-for-byte (re-capture the golden when a schema change is intentional), andToolSchemaParityTestsasserts every parameter the handler reads is declared in the schema. -
Build and publish:
- Add both test levels — a unit test of the
RoslynServicemethod against a deterministic fixture, AND a wire test through the MCP dispatcher (intests/SharpLensMcp.Tests/Mcp/) with exact value locks plus an error path. This is non-negotiable; see Testing.
Testing
Every test must satisfy the Testing Charter (C1–C9) in tests/SharpLensMcp.Tests/TESTING.md — the standing contract. The headline rules:
- Lock exact values (C1): assert a concrete name / count / substring / error code /
(line, column)— neverNotBeNull/> 0/ a type-only check as the sole assertion. - Both levels per tool (C4): a unit test against a
Fixtures/*.csfixture and a dispatcher (wire) test that unwrapscontent[0].text, plus an error path. - Right casing (C3): in-process Newtonsoft yields PascalCase
error.Code/meta.TotalCount; the MCP wire yields camelCase. Read the casing your test's path actually produces. - Deterministic (C7): the suite is serialized via
xunit.runner.json; fixture mutators always restore; the timeout test uses a forced-cancellation seam, not a timing race. - Out-of-process spine (C6):
StdioIntegrationTestsvalue-pins one tool per category over the real binary and runs thetools/listgolden over the stdio pipe. - Pre-commit gate (C9): build-clean + green is necessary but not sufficient — re-read each changed test and confirm it fails on a wrong answer.
Run the suite:
Key Files
License
MIT - See LICENSE for details.
来源:README.md,提交 35c4196
工具
0版本历史
1- v1.5.3最新Sep 16, 2026
评分
已满足适用检查项的 50%
该评分衡量此收录条目的完整程度,依据注册中心条目、端点与公开来源计算得出。它不是对服务器本身的评价,也不表示服务器安全或当前可用:可连通性是带日期的检查结果,评分也不包含安装量或用户评价。
- 近 30 天内可连通25 分
请保持端点能够响应 MCP 握手;SourceWeft 每 30 天探测一次。
- 已声明认证方式10 分
请在注册中心条目中声明凭据请求头,或保持端点能够响应探测,以便判定其认证方式。
- 已列出工具15 分
请让服务器在无凭据时响应 tools/list,或发布可在沙箱中无凭据启动的软件包。
- 描述必需15 分
- 图标10 分
- 文档15 分
- 许可证10 分