
This server gives Claude deep reflection capabilities over .NET assemblies. It exposes 31+ tools for type introspection, member analysis, and dependency resolution. You can enumerate methods and properties with filtering and pagination, extract XML documentation, trace type hierarchies, find implementations of interfaces, and parse project files and deps.json. It's built for scenarios where you need accurate type information rather than hallucinated API signatures: analyzing unfamiliar codebases, generating code against existing types, or building tooling that needs precise metadata about constructors, generics, and attributes. Runs as a global dotnet tool over stdio and includes smart caching for performance.
Sherlock MCP for .NET is a comprehensive Model Context Protocol (MCP) server that provides deep introspection capabilities for .NET assemblies. It enables Language Learning Models (LLMs) to analyze and understand your .NET code with precision, delivering accurate and context-aware responses for complex development scenarios.
This tool is essential for developers who want to harness LLM capabilities for:
core profile whose other tools load on demand in groupsexplore_package, explain_type and who_calls prompts give one-click entry points into common analysis workflowsModelContextProtocol 2.1.0 (GA)2026-07-28, and negotiates down automatically for clients on earlier revisionsget_member_source returns a member's original source through its PDB or Source Link, decompile_member / decompile_type decompile to C#, and compare_api_surface diffs two assembly or package versions and flags breaking changes.find_endpoints, find_service_registrations, find_ef_entities and find_handlers list ASP.NET Core endpoints, DI registrations, EF Core entities and MediatR handlers.project.assets.json (get_package_graph shows the restored graph), and project tools evaluate projects with MSBuild.open_assembly handles, workflow prompts (explore_package, explain_type, who_calls), Mermaid diagrams for hierarchies and call graphs, and on-demand tool groups with discovery hints under the core profile.CHANGELOG.md for full details.With the .NET 10 SDK, dnx downloads the package from NuGet and runs it directly — no install step:
dnx -v q --yes Sherlock.MCP.Server@2.15.0
--yes skips the interactive confirmation prompt, which an MCP client launching the server over stdio can't answer. -v q stops dnx from printing a "Skipping NuGet package signature verification." notice to stdout the first time it downloads a version, which would otherwise corrupt the MCP stream and fail that first connection. Both are dnx options, so they must come before the package id; anything after it is passed to the server. Pinning the version keeps launches reproducible; bump it when you want to upgrade. The package is published with the McpServer package type, so it is also listed as an MCP server on NuGet.org.
On any supported SDK (.NET 8 or later), install the global tool from NuGet (adds sherlock-mcp to your PATH):
dotnet tool install -g Sherlock.MCP.Server
Alternatively, during development you can run the server locally:
dotnet run --project src/server/Sherlock.MCP.Server.csproj
Sherlock runs as a standard MCP server that communicates over stdio.
This repository is also a Claude Code plugin marketplace. The sherlock plugin bundles the server (launched with dnx, so it needs the .NET 10 SDK) and a skill that teaches the agent the locate → orient → drill-in → relationships workflow:
/plugin marketplace add jcucci/dotnet-sherlock-mcp
/plugin install sherlock@dotnet-sherlock-mcp
The plugin starts the server with the core tool profile, which matches the tools the skill covers. The agent can add the other tools by group with load_tools when it needs them (see Tool groups). To expose every tool from the start, set SHERLOCK_TOOL_PROFILE=full in the environment Claude Code is launched from.
Each user runs these commands once; restart Claude Code (or run /reload-plugins) afterwards. To pick up a new release, run /plugin marketplace update dotnet-sherlock-mcp. If you previously registered Sherlock with claude mcp add, remove that entry (claude mcp remove sherlock) so the tools aren't loaded twice.
To offer the plugin to everyone working in a repository, commit the following to that repository's .claude/settings.json. Claude Code prompts each teammate to install it when they trust the folder, so nobody has to type the commands above:
{
"extraKnownMarketplaces": {
"dotnet-sherlock-mcp": {
"source": {
"source": "github",
"repo": "jcucci/dotnet-sherlock-mcp"
}
}
},
"enabledPlugins": {
"sherlock@dotnet-sherlock-mcp": true
}
}
Teammates still need the .NET 10 SDK on their PATH for dnx.
claude mcp add sherlock -- dnx -v q --yes Sherlock.MCP.Server@2.15.0
.vscode/mcp.json):{
"servers": {
"sherlock": {
"type": "stdio",
"command": "dnx",
"args": ["-v", "q", "--yes", "Sherlock.MCP.Server@2.15.0"]
}
}
}
sherlock-mcpsherlock-mcp command. Example JSON entry (refer to your client’s docs for exact file location/format):{
"servers": {
"sherlock": {
"command": "sherlock-mcp"
}
}
}
No arguments are required. The server self-registers all tools when launched.
Large tool lists cost agents context and discoverability (Claude Code switches to deferred Tool Search once tool descriptions grow past roughly 10% of the context window). Sherlock can start with a smaller surface:
| Profile | Tools | Contents |
|---|---|---|
full (default) | 47 | Every tool, including the deprecated per-kind member tools |
core | 24 | load_tools and invoke_tool (see Tool groups), discovery (find_assembly_by_class_name, find_assembly_by_file_name, find_assembly_by_nuget_package, get_project_output_paths, open_assembly), orientation (get_assembly_info, get_types_from_assembly, get_type_info, get_type_hierarchy), members and docs (get_type_members, search_members, analyze_method, get_xml_docs_for_type, get_xml_docs_for_member) and relationships (find_implementations_of, find_methods_returning, find_extension_methods_for, find_references_to, get_method_calls), source (get_member_source, decompile_member) and API diffs (compare_api_surface) |
Select a profile with the --profile argument or the SHERLOCK_TOOL_PROFILE environment variable (the argument wins). An unknown profile name stops the server with an error.
{
"servers": {
"sherlock": {
"command": "sherlock-mcp",
"args": ["--profile", "core"]
// or: "env": { "SHERLOCK_TOOL_PROFILE": "core" }
}
}
}
Under the core profile every other tool belongs to a group that the agent can load when it needs it:
| Group | Tools |
|---|---|
frameworks | find_endpoints, find_service_registrations, find_ef_entities, find_handlers |
project | analyze_solution, analyze_project, resolve_package_references, get_package_graph, find_deps_json_dependencies |
metadata | analyze_assembly, get_generic_type_info, get_nested_types, get_type_attributes, get_member_attributes, get_parameter_attributes |
decompile | decompile_type |
config | get_runtime_options, update_runtime_options |
legacy | the deprecated analyze_type, get_all_type_members and get_type_{methods,properties,fields,events,constructors} |
load_tools: with no arguments, lists every group, its tools and whether it is loaded. With groups=['frameworks', …], it adds those tools to the server's tool list and returns each loaded tool's name, description and input schema. The server then sends notifications/tools/list_changed. Under protocol revision 2026-07-28 that only reaches clients that asked for it through subscriptions/listen.
invoke_tool: calls a loaded tool by name with its arguments object and returns that tool's own result. Clients that never refresh their tool list can still use loaded groups this way. Calling a tool whose group isn't loaded fails with ToolNotLoaded and names the group.
Discovery hints: core results can carry hints[], each { tool, reason, group }, pointing at a related tool. Examples:
get_type_info on a type with nested types or generic parameters, or one that is a DbContext, controller or MediatR handler;find_implementations_of / find_references_to / find_extension_methods_for on those framework types or IServiceCollection;decompile_member → decompile_type;get_project_output_paths → analyze_project and get_package_graph.Dependency errors suggest get_package_graph. group is null for core tools.
Tool list caching: under core, tools/list is marked cacheScope: private, since a session's tool list grows as groups load. Under full, it stays public.
The full profile lists every tool and omits load_tools and invoke_tool.
You usually don't need to paste anything. Sherlock ships its usage guidance in the MCP instructions field returned at initialize, and most MCP clients (including Claude Code) surface that to the agent automatically — so the guidance stays correct and versioned with the package, with no copy-paste to maintain.
The snippets below are optional reinforcement. Keep them short and principle-based rather than enumerating tool names and workflows: a static list pasted into your repo will drift as Sherlock's tools evolve, whereas the tools' own descriptions (and the server instructions) always match the version you're running.
Tool names are exposed in
snake_case(get_type_members,search_members, …); argument names stay camelCase (projection,nameContains).
If you installed the Claude Code plugin, its skill already covers this. Otherwise, you can add a short, optional pointer to your project's CLAUDE.md:
## .NET Assembly Analysis
Use the Sherlock MCP tools (`get_type_members`, `search_members`, …) for .NET type/assembly
questions instead of guessing. Locate DLLs with the `find_assembly_by_*` / `get_project_output_paths`
tools rather than hardcoding bin paths. Start lean — `search_members` or `get_types_from_assembly`,
then drill in — and pass `projection='full'` only when you need parameters/attributes/modifiers.
The tools' own descriptions cover the specifics.
The single-file .cursorrules format is deprecated (and silently ignored in Cursor's Agent mode).
Add a Project Rule at .cursor/rules/sherlock.mdc instead:
---
description: Use Sherlock MCP for .NET assembly/type analysis
alwaysApply: true
---
- Prefer the Sherlock MCP tools (snake_case, e.g. `get_type_members`, `search_members`) over guessing about .NET APIs.
- Find DLLs with `find_assembly_by_*` / `get_project_output_paths`; don't hardcode `bin/Debug/<tfm>/*.dll`.
- Start lean (`search_members` / `get_types_from_assembly`); request `projection='full'` only when you need parameters/attributes/modifiers.
For tools that follow the cross-editor AGENTS.md convention, the same short pointer works — drop the Claude Code snippet above into your AGENTS.md.
For system-wide usage, add to your global agent settings:
For .NET work, use the Sherlock MCP tools (snake_case) to analyze assemblies, types, and members instead of guessing. Start lean and opt into projection='full' only when you need detail.
Below are compact prompt snippets you can paste into your chat to get productive fast. Adjust paths to your local DLLs.
General setup
You have access to an MCP server named "sherlock" that can analyze .NET assemblies. Prefer these tools for .NET questions and include short reasoning for which tool you chose. Ask me for the assembly path if missing.
Enumerate members for a type
Analyze: /absolute/path/to/MyLib/bin/Debug/net9.0/MyLib.dll
Type: MyNamespace.MyType
List methods, including non-public, filter name contains "Async", include attributes, return JSON.
Get XML docs for a member
Use GetXmlDocsForMember on /abs/path/MyLib.dll, type MyNamespace.MyType, member TryParse. Summarize the summary + params.
Find types and drill in
List types from /abs/path/MyLib.dll; then get type info for the first result and list its nested types.
Tune paging and filters
Use GetTypeMembers on /abs/path/MyLib.dll, type MyNamespace.MyType, kinds method, sortBy name, sortOrder asc, skip 0, take 25, hasAttributeContains Obsolete.
Browse lean, then get detail (projection)
On /abs/path/MyLib.dll, run GetTypeMembers for MyNamespace.MyType with the default summary projection to see signatures. Then re-call GetTypeMembers with kinds method, nameContains and projection='full' only for the methods I name to get their parameters and attributes.
Trace relationships and call sites
On /abs/path/MyLib.dll: FindImplementationsOf MyNamespace.IMyService. Then FindReferencesTo that interface with analysisDepth='il' to find callers, and GetMethodCalls on the most relevant method to see what it invokes.
Tool names: MCP clients call these tools in
snake_case—GetTypeMembers→get_type_members,SearchMembers→search_members, and so on. The PascalCase names used throughout this README match the underlying C# methods and the tool descriptions your client displays.
OpenAssembly: Returns a short asm_… handle to pass as assemblyHandle instead of assemblyPath on any tool that takes an assembly. The handle carries additionalAssemblies too, persists across server restarts (in handles.json under SHERLOCK_STATE_DIR, default the user's local app-data sherlock folder) and is pinned to the current build: after a rebuild, calls fail with StaleAssemblyHandle until it is reopenedAnalyzeAssembly: Complete assembly overview with public types and metadataGetAssemblyInfo: Assembly-level metadata — identity/version, target framework, and referenced assemblies (projection=full adds all assembly attributes); frameworkResolution reports what framework types were resolved against: the target framework's reference pack (referencePack, e.g. Microsoft.NETCore.App.Ref/8.0.x/ref/net8.0, plus Microsoft.AspNetCore.App.Ref and other packs named by the project's FrameworkReferences or the app's runtimeconfig.json), the assembly's own folder for self-contained output (appLocal), or the server's runtime when no matching pack is installed or the target is .NET Framework (hostRuntime, with the unmatched frameworks in missingFrameworks)FindAssemblyByClassName: Locate assemblies declaring a public type by simple, full or nested name; skips obj/, ref/, refint/, node_modules/, packages/, TestResults/ and dot-directories, and ranks bin/ matches firstFindAssemblyByFileName: Find assemblies by file name under a root directory, with the same exclusions and rankingFindAssemblyByNugetPackage: Resolve a DLL from the local NuGet cache by package id (optional version/tfm)GetTypesFromAssembly: List all public types with metadata (paginated)AnalyzeType (deprecated): Type metadata plus all members; use GetTypeInfo + GetTypeMembers projection=fullGetTypeInfo: Detailed type metadata (accessibility, generics, nested types)GetTypeHierarchy: Inheritance chain and interface implementations; format='mermaid' returns a Mermaid classDiagramGetGenericTypeInfo: Generic parameters, arguments, and variance informationGetTypeAttributes: Custom attributes declared on typesGetNestedTypes: Nested type declarationsGetTypeMembers: Methods, properties, fields, events and constructors of a type in one paginated list. Narrow with kinds (method|property|field|event|constructor), nameContains and hasAttributeContains; summary items are { kind, name, signature }, projection=full adds each kind's structured fieldsAnalyzeMethod: Deep method analysis with overloads and attributesGetTypeMethods, GetTypeProperties, GetTypeFields, GetTypeEvents, GetTypeConstructors (use GetTypeMembers with kinds) and GetAllTypeMembers (use GetTypeMembers projection=full)SearchMembers: Search a whole assembly for members whose name contains a fragment — the entry point when you know a member name but not its declaring type. Filter by memberKinds (method|property|field|event|type).FindImplementationsOf: Types implementing an interface or deriving from a base class (open-generic match supported)FindMethodsReturning: Methods whose return type matches a given type (open-generic match supported)FindExtensionMethodsFor: Extension methods that extend a given type (scans static classes by this-parameter)FindReferencesTo: Broader sweep across parameters, fields, properties, events, and generic arguments; pass analysisDepth='il' to also resolve inbound callers from method bodiesIn the frameworks tool group (always listed under full). Each tool scans one or more assemblies (additionalAssemblies) and pages with maxItems / continuationToken. The default summary projection is lean, and projection='full' adds the assembly path and detail fields.
FindEndpoints: ASP.NET Core endpoints. Controller actions are read from [Route] / [Http*] / [AcceptVerbs] attribute routes, with [controller] / [action] substituted. Minimal APIs come from Map{Get,Post,Put,Delete,Patch,Methods,Group,Fallback} calls read from IL. Filter with routeContains and httpMethod. Minimal-API routes and handlers are taken from the string literal and method reference next to each call, so a route held in a variable or field comes back as null, and MapGroup prefixes are listed as separate minimalApiGroup entries rather than combinedFindServiceRegistrations: Microsoft.Extensions.DependencyInjection registrations, read from IL calls to Add{Singleton,Scoped,Transient}, TryAdd*, AddKeyed*, AddHostedService and the ServiceDescriptor factories. Lifetime, service and implementation come from generic arguments or typeof(...) operands, and factory and instance registrations report implementation: null. Filter with serviceType, implementationType and lifetimeFindEfEntities: The DbSet<T> properties, declared or inherited, of every DbContext subclass. Filter with contextType and entityTypeFindHandlers: MediatR-style handlers: concrete implementations of IRequestHandler<,> / IRequestHandler<>, INotificationHandler<>, IStreamRequestHandler<,> and IPipelineBehavior<,> from the MediatR or Mediator namespaces. Filter with messageType and kindGetMethodCalls: Read a method's IL body to list what it calls and which fields it touches — the "what does this method do?" question signature-level tools can't answer (aggregates across overloads; use .ctor/.cctor for constructors). format='mermaid' returns a Mermaid flowchart call graph, and depth (up to 5) follows callees defined in the same assemblyCompareApiSurface: Diff the public API of two versions of an assembly. left (old) and right (new) each take an assembly path, an asm_… handle, or a NuGet package as Package.Id@1.2.3 resolved from the local cache (omit the version for the highest cached one; tfm picks the target framework; by default both sides use one target framework both packages ship, with a warning when they share none). Public types and members, plus protected ones on inheritable types, are compared. Breaking changes are flagged with reasons: removed types or members, changed return/field/property types, reduced accessibility, classes that became sealed, abstract or static, removed base types or interfaces, new generic constraints, abstract members added to interfaces or inheritable classes, members that became static/instance, abstract or non-virtual, fields that became readonly, removed or less accessible property accessors, setters that became init, changed parameter modifiers, parameters that are no longer optional or params, and changed enum values. Additions and compatible changes (parameter renames, new default values, changed constants, new params, members moved to a base type) are listed without the flag. The default summary returns counts and the breaking changes; projection='full' lists every change with leftSignature, rightSignature and reasons[] (narrow with breakingOnly and namespaceFilter). Paged with maxItems / continuationTokenGetMemberSource: The member's original source, with comments and real names, read through the assembly's portable PDB (embedded in the DLL or beside it). It uses source embedded in the PDB, then the local file the assembly was built from, then the file at the PDB's Source Link URL. Each candidate, embedded source included, must match the SHA-1/SHA-256 checksum the PDB recorded (CRLF/LF differences are tolerated); otherwise that overload falls back to decompiled C#. Local files are read only from absolute, non-UNC paths and only up to 5 MB. The result's origin is embedded, local, sourcelink, decompiled or mixed, each overload reports its document, url and lines, and note explains any fallback. The slice covers the member's /// docs, attributes, signature and bodyDecompileMember: Decompile one member (method, property, field, event or constructor) to C# with ICSharpCode.Decompiler. Returns every overload of the name, or one overload when parameterTypes is given (e.g. string,int; an empty string selects the parameterless overload). Use .ctor/.cctor for constructorsDecompileType: Decompile a whole type to C#. In the decompile tool group; prefer DecompileMemberAll three tools page their source by line: maxLines (default 400, max 5000) caps the page size, a page also stops early once it reaches about 90,000 characters so it always fits the response limit, and each page reports startLine, lineCount, totalLines, truncated and a continuationToken for the next page. Lines over 2,000 characters are clipped with a /* … more characters clipped */ marker and counted in clippedLines. Pass additionalAssemblies (or use an assemblyHandle opened with them) when dependencies live outside the assembly's folder; their folders are searched when resolving referenced types and their file stamps are part of the cache key. The full decompilation is cached by file stamp, so later pages are cheap. A type the assembly only forwards (e.g. System.String in a System.Runtime.dll facade) returns TypeForwarded with the defining assembly in recommendedParams.
GetMemberAttributes: Attributes for specific membersGetParameterAttributes: Parameter-level attribute informationGetXmlDocsForType: Extract type-level XML documentationGetXmlDocsForMember: Member-specific documentation (summary/params/returns/remarks)AnalyzeSolution: Parse .sln files and enumerate projectsAnalyzeProject: Project metadata, references, and build configurationGetProjectOutputPaths: Resolve output directories for different configurationsResolvePackageReferences: Map NuGet packages to cached assembliesGetPackageGraph: Restored NuGet graph from obj/project.assets.json (exact direct and transitive versions per target framework and RID)FindDepsJsonDependencies: Parse deps.json for runtime dependenciesProject tools evaluate the project with the installed .NET SDK (dotnet msbuild -getProperty/-getItem, evaluation only, no build), so Directory.Build.props/targets, central package management, conditions, imports and computed output paths are honoured. The SDK is chosen by the project's global.json. When no SDK is available (for example in the container image) or evaluation fails, they fall back to parsing the project file's XML, and evaluation.mode / evaluation.reason say so. Evaluation runs the repository's .props/.targets and SDK resolvers; set SHERLOCK_PROJECT_EVALUATION=off or call update_runtime_options projectEvaluation=off to use only the XML parser.
GetRuntimeOptions: Current server configuration and defaultsUpdateRuntimeOptions: Modify pagination, caching, search, Source Link fetch and project evaluation behaviorget_member_source fetches Source Link files over HTTPS only from known source hosts by default (raw.githubusercontent.com, gitlab.com, bitbucket.org, api.bitbucket.org, dev.azure.com, *.visualstudio.com), because the URL comes from the PDB and could otherwise point anywhere. Set SHERLOCK_SOURCE_FETCH to off, known-hosts (default) or any, or change it at runtime with update_runtime_options sourceFetch=… and addSourceFetchHosts / removeSourceFetchHosts. A host gets 5 seconds to respond and 30 seconds to deliver the file, files over 5 MB are refused, redirects are followed manually (at most 3) and each hop must pass the same HTTPS and host checks, a host that can't be connected to is not retried for the rest of the session, a fallback caused by a network failure is not cached, and verified files are cached under SHERLOCK_STATE_DIR/sources.
Three resource templates let clients fetch a single type, doc entry or cached package without another tool call. Every variable is percent-encoded.
sherlock://assembly/{path}/type/{fullName}: Type metadata, the same payload as get_type_infosherlock://assembly/{path}/docs/{memberId}: XML documentation for a documentation id such as T:Ns.Type or M:Ns.Type.Method(System.String)sherlock://nuget/{packageId}/{version}: The assembly a cached NuGet package version resolves to, the same payload as find_assembly_by_nuget_packagesearch_members, get_types_from_assembly and the find_* reverse-lookup tools return a resource_link to the type resource for each distinct type on the page, after the JSON text block. resources/read carries private caching hints; a URI whose assembly, type, doc id or package doesn't exist fails with -32602.
Template variables support completion/complete, returning at most 100 values:
path: recently loaded assemblies, then .dll/.exe files and subfolders of the folder being typedfullName: type names (metadata form, e.g. List`1) from the assembly in the path context argumentmemberId: documentation ids from the XML file next to the path assemblypackageId / version: package folders and their versions (newest first) in the local NuGet cacheThree prompts encode common tool sequences as one-click entry points (Claude Code lists them as /mcp__sherlock__<name> slash commands). Each returns a single user message that walks the agent through the tools; they only call tools in the core profile, so they work under either profile.
explore_package (packageId, version?): Resolves the package from the local NuGet cache (highest cached version when version is omitted), orients on the assembly, and summarizes its main types and entry pointsexplain_type (assemblyPath, typeName): Hierarchy, members, XML docs and usages of one typewho_calls (assemblyPath, typeName, memberName, additionalAssemblies?): Inbound callers of a member from IL, optionally across a comma-separated list of other assembliesassemblyPath, typeName (with assemblyPath in the completion context), packageId and version (with packageId in the context) support completion/complete the same way as the resource template variables. prompts/list carries the same public caching hints as tools/list.
All member analysis tools support comprehensive filtering and pagination:
Filtering Options:
caseSensitive (bool): Case-sensitive type/member matchingnameContains (string): Filter by member name substringhasAttributeContains (string): Filter by attribute type substringincludePublic / includeNonPublic (bool): Visibility filteringincludeStatic / includeInstance (bool): Member type filteringPagination:
skip / take (int): Standard offset paginationmaxItems (int): Maximum results per request (default 50; FindReferencesTo defaults to 25)continuationToken (string): Token-based pagination for large datasetssortBy / sortOrder (string): Sort by name/access in asc/desc orderMost enumerating tools default to a lean summary projection and let you opt into the heavier full payload only when you need it. Reach for full deliberately — summary is usually enough to decide your next call.
projection (summary | full): supported by GetTypesFromAssembly, GetTypeMembers, GetTypeMethods, GetAssemblyInfo, GetMethodCalls, FindImplementationsOf, FindMethodsReturning, FindExtensionMethodsFor, FindReferencesTo, FindEndpoints, FindServiceRegistrations, FindEfEntities, and FindHandlers. summary returns just enough to browse (e.g. { kind, name, signature } for members); full adds structured fields (parameters, attributes, return type, modifiers, etc.). Note: the deprecated GetTypeProperties/Fields/Events/Constructors have a single fixed shape and take no projection.analysisDepth (signatures | il): FindReferencesTo only. signatures (default) scans member declarations; il additionally scans method bodies for inbound callers (slower).format (json | mermaid): supported by GetTypeHierarchy (classDiagram) and GetMethodCalls (flowchart, with depth for transitive calls). Mermaid results return { diagram, nodeCount, edgeCount, truncated, note }, bounded by maxNodes (default 50, max 200).additionalAssemblies (string[]): widen the search scope for GetTypeHierarchy and the reverse-lookup tools. GetTypeHierarchy.derivedTypes stays null until you pass this.noCache (bool): bypass the response cache for a single call when you suspect stale results. Every tool that reads an assembly or project caches its response, keyed by the file stamps of its inputs (the assembly and any additionalAssemblies, the project.assets.json of the project whose bin output holds the assembly, the app's runtimeconfig.json, the XML doc file beside it, or the project file, obj/project.assets.json and any Directory.Build.props, Directory.Build.targets, Directory.Packages.props or global.json above it), so a rebuild or restore invalidates it automatically. The find_assembly_by_* lookups and resolve_package_references (which read the shared NuGet package cache), configuration and handle tools are not cached. Two cases the stamps don't catch, where noCache=true is the fix: a dependency DLL dropped in beside an assembly that previously resolved only partially (keys stamp the assembly, not its siblings), and a restore for a project whose project.assets.json lives outside obj/ or artifacts/obj/<project>/ (e.g. a custom BaseIntermediateOutputPath), and a targeting pack installed after an assembly was first resolved against the host runtime (a noCache call notices the new pack and rebuilds the assembly's inspection context).assemblyHandle (string): accepted by every tool that takes assemblyPath, as an alternative to it (pass one, not both). Get one from open_assembly; it also supplies the additionalAssemblies it was opened with, and any you pass explicitly are added to them.Type Resolution:
Namespace.Type), simple names (Type), and nested types (Outer+Inner)caseSensitive parameterAll tools return a stable JSON envelope:
{ "kind": "type.list|member.methods|...", "version": "1.0.0", "data": { /* result */ }, "hints": [ /* optional */ ] }
The envelope is serialized as compact (unindented) JSON in the tool's text content block. The core browsing tools also advertise an MCP outputSchema and return the same envelope as structuredContent, so clients can validate and consume results without parsing text: open_assembly, search_members, get_types_from_assembly, get_type_info, get_type_members, get_type_methods, get_assembly_info, get_method_calls, decompile_member, get_member_source, compare_api_surface, find_implementations_of, find_methods_returning, find_extension_methods_for and find_references_to. Their schemas describe the default summary projection; projection='full' items add fields on top of it. Error results never carry structuredContent.
Error results are flagged with MCP's isError: true, so clients can tell a failure from a result without parsing the text. Errors use a consistent shape. Every error carries kind, version, code, and message; some add details, and guided errors add a suggestion, alternativeTools, or recommendedParams to point the agent at a next step:
{
"kind": "error",
"version": "1.0.0",
"code": "MethodNotFound",
"message": "No method named 'Parse' was found on type 'MyApp.Config' in MyApp.dll.",
"suggestion": "Verify the type and method names. Use get_type_members with kinds=method to list available methods, or set includeNonPublic=true for private methods.",
"alternativeTools": ["get_type_members", "analyze_method"]
}
Error codes:
AssemblyNotFound (recommendedParams.similarFiles lists near-miss assemblies in the same folder), TypeNotFound / MemberNotFound (recommendedParams.candidates lists the closest names), MethodNotFound, PackageNotFound, VersionNotFound, XmlNotFound, ProjectNotFound, FileNotFound, OverloadNotFound, TypeForwarded (recommendedParams names the defining assembly), AssetsFileNotFound, TargetFrameworkNotFound, ToolNotFoundAmbiguousTypeName (only for clients that can't elicit; recommendedParams.candidates lists the matching full names), AmbiguousTargetFramework, InvalidArgument, InvalidProjection, InvalidAnalysisDepth, InvalidFormat, InvalidContinuationTokenStaleAssemblyHandle, UnknownAssemblyHandle, ToolNotLoaded (the error names the group to load)InvalidAssembly (not a managed assembly), DependencyNotFound (a referenced assembly couldn't be loaded), DependencyResolutionFailed, InvalidAssetsFile, AccessDeniedResponseTooLarge, InternalErrorShipped features and planned work are summarized in src/docs/roadmap.md; the live checklist is tracked in #87.
Contributions are welcome. This repo includes an .editorconfig with modern C# preferences (file-scoped namespaces, expression-bodied members, 4-space indentation).
This project uses Conventional Commits for automated changelog generation. All commits must follow this format:
type(scope): description
Valid types:
feat - A new featurefix - A bug fixdocs - Documentation only changesstyle - Code style changes (formatting, semicolons, etc)refactor - Code change that neither fixes a bug nor adds a featureperf - Performance improvementtest - Adding or correcting testsbuild - Changes to build system or dependenciesci - Changes to CI configurationchore - Other changes that don't modify src or test filesrevert - Reverts a previous commitExamples:
git commit -m "feat(tools): add new assembly analysis tool"
git commit -m "fix: resolve null reference in type loader"
git commit -m "docs(readme): update installation instructions"
# Restore .NET tools (versionize, husky)
dotnet tool restore
# Install git hooks for commit validation
dotnet husky install
dotnet build and dotnet test locally before opening a PR.Releases are cut from a pull request that bumps every version pin to the new version: <Version> in src/server/Sherlock.MCP.Server.csproj, both version fields in server.json, the Claude Code plugin's version in plugins/sherlock/.claude-plugin/plugin.json and .claude-plugin/marketplace.json, the pinned dnx version in plugins/sherlock/.mcp.json and in this README, and the CHANGELOG.md heading ([Unreleased] → [x.y.z] - date). server.json is packed into the NuGet package as .mcp/server.json. Don't run dotnet versionize on top of that; it would bump the version again.
After the PR merges, tag the merge commit on main and push the tag:
git tag -a vX.Y.Z -m "Release X.Y.Z"
git push origin vX.Y.Z
The release workflow will automatically:
server.json, the project version and the Claude Code plugin versions all matchmcp-name: io.github.jcucci/dotnet-sherlock-mcp
Sherlock MCP for .NET is licensed under the MIT License.