RoselineMCP

Roslyn code intelligence for AI coding agents, over MCP. Give Claude, Cursor, and Copilot a semantic view of your C# solution — symbols, references, call graphs, surgical edits — so they navigate by structure instead of re-reading source. Measured 85% fewer tokens (median) →

📖 Documentation, tool reference & the honest benchmark →

Why RoselineMCP
Quick Start
Features
Tech Stack
Getting Started
MCP Client Compatibility
Available Tools
Tool Annotations
Tool Compatibility Policy
Architecture
Project Structure
Compile guard
Security
Documentation
Roadmap
Contributing
License
Acknowledgments

Why RoselineMCP

Your coding agent shouldn't read a 700-line file to change one method. Source code dominates an
agent's token budget, so the cheapest win is to stop feeding it whole files.

RoselineMCP wraps the Roslyn compiler platform as an MCP server.
Instead of dumping source into the model, it answers structural questions precisely — where is
this symbol used, what implements this interface, who calls this method, what's the shape of this
file — and it edits surgically: a member-level diff, not a whole-file rewrite.

On RoselineMCP's own source, the read-only navigation tools returned a median 85% fewer tokens
per task (pooled, size-weighted: 93%) than reading the corresponding files —
measured honestly, weak cases included.

search_symbols on Program.cs: 2,093 tokens → 120 (−94%). The agent gets the shape of the
file; you skip the wall.

Quick Start

Any MCP client that speaks dnx (the .NET equivalent of npx) runs it on demand — no install step.
Requires the .NET 10 SDK.

// claude_desktop_config.json  ·  .vscode/mcp.json  ·  ~/.cursor/mcp.json
{
  "mcpServers": {
    "roseline": { "command": "dnx", "args": ["RoselineMCP", "--yes"] }
  }
}

Then ask your agent to "find every caller of OrderService.Checkout" or "rename Foo to Bar
across the solution." Prefer a pinned NuGet install or Docker? See
Getting Started.

Features

Token-efficient code navigation -- symbols, references, call graphs, type hierarchies, and file outlines via Roslyn instead of whole files. A measured 85% median token reduction per task (93% pooled, size-weighted) -- see the benchmark.
Surgical code edits -- replace/add/delete a member or rename a symbol solution-wide, emitting a unified diff instead of a whole-file rewrite. Preview by default.
Comprehensive analysis & auto-fix -- diagnostics across a solution (Roslyn + Roslynator) with automated fixes and reviewable patches.
Read-only by default -- the seven navigation tools and the diagnostics/patch tools never touch disk; the three write tools require an explicit previewOnly: false.
Compile guard (opt-in) -- a PostToolUse hook that puts the compiler's verdict behind every file write, not just RoselineMCP's own -- see Compile guard below.
Works with your client -- Claude Desktop, VS Code (Copilot / MCP), Cursor. Install via dnx, NuGet global tool, or Docker.
Honest, reproducible benchmark -- run it against your own solution: dotnet run --project RoselineMCP.TokenBenchmark -c Release.

Tech Stack

Layer	Technology
Runtime	.NET 10.0
Compiler Platform	Roslyn (Microsoft.CodeAnalysis) 5.6.0
Analyzers	Roslynator 4.15.0
MCP SDK	ModelContextProtocol 2.2.0
Diff Engine	DiffPlex 1.9.0
Build System	MSBuild 18.8.2
Hosting	Microsoft.Extensions.Hosting 10.0.10

Versions above are kept in sync with RoselineMCP/RoselineMCP.csproj — that file is the source of truth if this table ever drifts.

Getting Started

Prerequisites

NuGet global tool: .NET 10.0 SDK or later
Docker: Docker Desktop or Docker Engine
Build from source: .NET 10.0 SDK + MSBuild (included with Visual Studio or .NET SDK)
MCP client: Claude Desktop or any MCP-compatible client

Installation

Claude Desktop, one click: download RoselineMCP.mcpb from the
latest release and open it —
Claude Desktop shows an install dialog, no config editing. (It launches via dnx under the hood,
so the .NET 10 SDK is still required.) Prefer to edit config yourself, or using another client?
Use one of the options below.

Option 1 -- dnx (no install step) (recommended)

RoselineMCP ships an MCP server registry manifest, so any MCP client that
understands the dnx launcher (the .NET equivalent of npx — resolves and runs a NuGet-packaged
tool on demand, without a separate dotnet tool install step) can start it directly. Requires the
.NET 10.0 SDK.

{
  "mcpServers": {
    "roseline": {
      "command": "dnx",
      "args": ["RoselineMCP", "--yes"]
    }
  }
}

Add this to your Claude Desktop or VS Code MCP configuration (see
MCP Client Compatibility below for exact file locations per client).
dnx downloads and caches the tool on first use, so there's nothing to pre-install globally.

Option 2 -- NuGet Global Tool (offline / pinned-version installs)

Requires .NET 10.0 SDK or later.

dotnet tool install -g RoselineMCP

After installation, the roseline-mcp command is available globally.

Claude Desktop configuration (NuGet global tool)

Add to your Claude Desktop configuration file (claude_desktop_config.json):

{
  "mcpServers": {
    "roseline": {
      "command": "roseline-mcp"
    }
  }
}

Config file location:

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json

Windows: %APPDATA%\Claude\claude_desktop_config.json

Option 3 -- Docker

No SDK required. Works on any platform with Docker installed.

docker run -i --rm phmatray/roseline-mcp:latest

Claude Desktop configuration (Docker)

{
  "mcpServers": {
    "roseline": {
      "command": "docker",
      "args": [
        "run",
        "-i",
        "--rm",
        "phmatray/roseline-mcp:latest"
      ]
    }
  }
}

Note: The -i flag is required for stdio transport. The --rm flag removes the container after the session ends.

Option 4 -- Build from Source

git clone https://github.com/Atypical-Consulting/RoselineMCP.git
cd RoselineMCP
dotnet build
dotnet test

Claude Desktop configuration (build from source)

{
  "mcpServers": {
    "roseline": {
      "command": "dotnet",
      "args": ["run", "--project", "/path/to/RoselineMCP/RoselineMCP.csproj"]
    }
  }
}

MCP Client Compatibility

RoselineMCP speaks plain stdio MCP, so it should work with any MCP-compatible client. The
snippets below are documented, not independently verified in every case — we've confirmed the
protocol-level behavior (stdio transport, tool discovery, JSON responses) works correctly, but we
have not personally exercised each client's own configuration UI/file end to end. If one of these
doesn't work as written for your client version, please open an issue.

Claude Desktop

Edit claude_desktop_config.json (see file locations under Installation
above) and add a roseline entry under mcpServers, using any of the four install options shown
above (dnx, global tool, Docker, or build-from-source).

VS Code (GitHub Copilot / MCP extension)

Add an entry to your workspace or user mcp.json (Command Palette → "MCP: Open User
Configuration", or .vscode/mcp.json in the workspace):

{
  "servers": {
    "roseline": {
      "command": "dnx",
      "args": ["RoselineMCP", "--yes"]
    }
  }
}

Substitute "command": "roseline-mcp" (no args) if you installed via the NuGet global tool
instead.

Cursor

Add a roseline entry to ~/.cursor/mcp.json (global) or .cursor/mcp.json (project-local),
using the same command/args shape as the VS Code snippet above.

Available Tools

Every tool's description — the text an MCP client shows the model at tool-selection time — states
its own Limitations: and shows one Example: call, within a test-enforced word ceiling; the
twelve tools with an optional project share one wording for the worktree caveat below, and
resolvedPath is how you confirm which checkout actually answered. See
docs/API.md § Tool description contract.

1. AnalyzeSolution

Analyzes an entire C# solution for diagnostics. Read-only — never modifies files on disk.
pathOrGit also accepts an http(s):// Git URL, which is shallow-cloned to a temp directory,
analyzed, and deleted afterward.

analyzeSolution({
  pathOrGit: "/path/to/solution.sln",
  include: "Core",              // Optional: only project names containing this substring
  exclude: "Test",              // Optional: skip project names containing this substring
  severity: "warning",          // Optional: minimum severity (Error|Warning|Info|Hidden)
  maxDiagnostics: 100           // Optional: Maximum diagnostics to return (default: 100)
})

Returns: solution file name, project count, a diagnosticSummary (counts by severity), a
topDiagnostics array with project/file/line/column/id/severity/message per diagnostic, and
analyzerLoad — present only when an analyzer reference contributed nothing (merged across the
analyzed projects), naming it and why.

2. ListDiagnostics

Gets detailed diagnostics for a specific project. Read-only — never modifies files on disk.
project is optional and accepts the same references as the navigation tools (name,
directory, .csproj, or .sln path); when omitted, the solution/project is auto-discovered from
the working directory.

listDiagnostics({
  project: "MyProject.csproj",     // Optional: name, directory, .csproj, or .sln; auto-discovered if omitted
  ids: ["CS0168", "CS0219"],       // Optional: Filter by diagnostic IDs
  files: ["Controller.cs"],        // Optional: substring match against each diagnostic's file path (case-insensitive; NOT a glob pattern)
  max: 50                          // Optional: Maximum results
})

Returns: project name, resolvedPath (the absolute .sln/.csproj actually loaded),
totalDiagnostics count, the filtered diagnostics list, stats (counts grouped by ID and by
severity), suggestedFixableIds — diagnostic IDs a code fix provider is actually registered for,
whether it ships with Roslyn, in the bundled Roslynator catalog, or inside one of the project's
own analyzer references — and analyzerLoad, which names every analyzer reference that
contributed nothing (and why), present only when there is something to say.

3. ApplyFixes

Applies automated code fixes for specified diagnostics. Defaults to preview mode:
previewOnly defaults to true, so calling this tool without setting it never writes to disk —
you must pass previewOnly: false explicitly to apply changes. project is optional and
accepts the same references as the navigation tools (name, directory, .csproj, or .sln path);
when omitted, the solution/project is auto-discovered from the working directory.

applyFixes({
  ids: ["CS0168", "RCS1001"],   // Diagnostic IDs to fix
  project: "MyProject.csproj",  // Optional: name, directory, .csproj, or .sln; auto-discovered if omitted
  previewOnly: false             // Optional (default: true). Set false to write changes to disk.
})

Returns: project name, resolvedPath (the absolute .sln/.csproj actually loaded),
fixedCount, fixersApplied (diagnostic IDs actually fixed), changedFiles
(relative to resolvedPath's directory, forward slashes — the same path base as the navigation
tools), a unified
diff patch, notes (the scope — which project was fixed and which of the solution's projects were
not analyzed, plus any linked file whose write reaches a sibling — and skipped/failed IDs and status
messages), previewOnly echoing back what the caller asked for, applied (whether anything
actually reached disk), verification — the compiler's verdict on the fixed code — and
analyzerLoad (present only when an analyzer reference contributed nothing, so "no diagnostics
found for X" can be told apart from "the analyzer that reports X never loaded"). Fixers are looked
up in the Roslyn built-ins, then the bundled Roslynator catalog, then the project's own analyzer
references; the bundled provider wins for an ID both carry.

A .sln target fixes one project — its primary project — and that scope is enforced on the
write path, not only announced by the confirmation prompt: only the anchor project's documents are
verified and written, and a caller who never sees a prompt (preview, a non-eliciting client, an
unattended host) still reads which projects were skipped in notes. Pass a .csproj to fix a
specific project — then nothing is reported as skipped, since that is what was asked for.

14. CheckCompilation

Answers "does this compile right now, and what broke" against on-disk state — the replacement
for a dotnet build round trip in an agent's edit loop. Read-only, and compiler diagnostics only:
analyzers cost several times a bare compile, which is the whole reason this tool is fast enough to
run after every edit.

It reports on whatever is on disk, whoever wrote it, so it works just as well for edits made by
other tools. The speed comes from the warm MSBuildWorkspace the server already holds: the first
call of a session pays a cold load, every call after it reuses an incremental Roslyn compilation.

checkCompilation({
  project: "MyApp.sln",  // Optional: name, directory, .csproj, or .sln; auto-discovered if omitted
  max: 20                 // Optional (default: 20). The rest are counted in `omitted`.
})

Returns: resolvedPath, compiles, errors (omitted when it compiles), omitted, scope
(the projects compiled), scopeComplete and notes.

check_compilation vs list_diagnostics: check_compilation answers "is it still
building?" — compiler errors, fast, for the edit loop. list_diagnostics answers "what should I
clean up?" — analyzer diagnostics, statistics, and which IDs are auto-fixable. Reach for the
first after an edit and the second when exploring.

4. CreatePatch

Generates a unified diff between two text versions. Read-only — operates purely on the provided
strings, never touches the filesystem.

createPatch({
  before: "original code",
  after: "modified code",
  fileName: "Example.cs",        // Optional: For display in diff
  ignoreWhitespace: false,       // Optional: ignore whitespace-only differences
  ignoreCase: false              // Optional: ignore case differences
})

Returns: the unified diff patch, hasChanges, linesAdded, linesRemoved, and the
fileName/summary used in the diff header.

Code Navigation Tools (read-only)

These tools return precise structure instead of whole files, so an AI agent can orient itself
in a codebase while spending far fewer tokens than reading source directly. All are read-only and
take an optional project (name, directory, .csproj path, or .sln path) — when omitted,
RoselineMCP auto-discovers the solution/project from its working directory. When the project belongs
to a solution, the whole solution is loaded and symbol search/resolution spans every project in it
(including sibling projects the requested project doesn't reference), so references/renames span
projects. Full request/response shapes are in docs/API.md.

Working in a git worktree? Auto-discovery is anchored to the server's working directory —
the one the MCP client launched RoselineMCP in — not yours. They differ whenever work happens in a
git worktree (e.g. .claude/worktrees/<name>), which sits below the discovery walk's reach, so an
omitted project resolves the main checkout instead. Every tool that takes an optional
project — the seven navigation tools, both edit tools, listDiagnostics and applyFixes —
reports resolvedPath, the absolute .sln/.csproj that actually answered — the .sln when the
solution was loaded and contains the project, otherwise the .csproj that was opened directly
(e.g. a project not listed in its nearest ancestor .sln). Check it, and pass
an absolute path as project to target a specific checkout. (analyzeSolution is the exception:
its pathOrGit is required, so it auto-discovers nothing.)

Failures report it too, which is where you will usually meet this: the wrong checkout answers
NotFoundError: Symbol not found: 'X' rather than a plausible-looking success. The failure
envelope's error.resolvedPath names the checkout that was searched, and is omitted entirely when
the call failed before resolving anything. See docs/API.md.

Relative file paths hang off resolvedPath. The navigation tools' file/definitionFile,
applyFixes/editMember/renameSymbol's changedFiles and patch headers, and
verification.errors[]/checkCompilation's errors[] file, are relative
to the directory containing resolvedPath — so dirname(resolvedPath) + <returned path> is the
real file, and a returned patch applies (git apply -p1) from that directory. That is the
solution root in the usual case, and the project's own directory whenever a .csproj answered
directly — including a project not listed in its nearest ancestor .sln. One exception:
listDiagnostics/analyzeSolution report file absolute, so there is nothing to join. See
docs/API.md.

Tool names on the wire are snake_case. The section headings below use friendly
PascalCase/camelCase for readability, but the actual MCP tool names returned by tools/list
(and expected by tools/call) are: search_symbols, get_symbol_info, find_references,
find_implementations, get_call_graph, get_type_hierarchy, get_symbol_at_position,
edit_member, rename_symbol
(matching the existing analyze_solution / list_diagnostics / apply_fixes /
check_compilation / create_patch).

5. SearchSymbols

Find symbols by wildcard/substring name pattern, or outline a single file.

searchSymbols({
  project: "MyApp.Core",
  query: "*Service",             // Substring, or wildcard with * and ? — omit to outline a file
  file: "UserService.cs",        // Optional: restrict to one file, or outline it when query omitted
  kinds: ["class", "method"],    // Optional: filter by kind (also accepts "type" / "member")
  max: 50                        // Optional (default: 50)
})

Returns: symbols (name, fullName, kind, signature, file, line — file is relative to resolvedPath's directory; the single-file outline instead returns name, kind, signature, line, containingType), totalFound, truncated (omitted when not capped).

6. GetSymbolInfo

The compact "go to definition": a symbol's declaration metadata and (optionally) its source.

getSymbolInfo({
  project: "MyApp.Core",
  symbol: "Acme.Users.UserService.GetUser",  // Simple or fully-qualified name
  includeSource: true                          // Optional (default: true)
})

Returns: name, fullName, kind, signature, and (each omitted when empty/absent) modifiers, baseTypes, interfaces, documentation, definitionFile/Line, and source. Accessibility is already part of signature; definitionFile is relative to resolvedPath's directory.

7. FindReferences

Every use site of a symbol across the solution, as location + one-line snippet.

findReferences({ project: "MyApp.Core", symbol: "GetUser", includeDefinition: false, max: 100 })

Returns: references (file — relative to resolvedPath's directory, line, snippet), totalReferences, truncated (omitted when not capped).

8. FindImplementations

Implementations of an interface/member, overrides of a virtual/abstract member, or derived types of a class.

findImplementations({ project: "MyApp.Core", symbol: "IRepository", max: 100 })

Returns: implementations (symbol summaries), totalFound, truncated.

9. GetCallGraph

A depth-bounded caller and/or callee graph for a method, with cycle detection.

getCallGraph({
  project: "MyApp.Core",
  method: "Handle",
  direction: "callers",   // "callers" (default) | "callees" | "both"
  depth: 1,               // 1-3 (default: 1)
  max: 50                 // Optional: nodes expanded per direction
})

Returns: callers/callees trees of nodes (fullName with simple parameter-type names, file — relative to resolvedPath's directory, line, truncated, children). Call GetSymbolInfo for a node's full signature.

10. GetTypeHierarchy

A type's base-class chain, implemented interfaces, and/or derived types.

getTypeHierarchy({
  project: "MyApp.Core",
  type: "SqlRepository",
  direction: "both",   // "base" | "derived" | "both" (default)
  max: 100             // Optional: maximum derived types returned (default: 100)
})

Returns: baseTypes, interfaces, derivedTypes (as symbol summaries).

13. GetSymbolAtPosition

The symbol living at a file:line(:column) position — turn a diagnostic, stack trace, or grep hit
into a symbol name without reading the file.

getSymbolAtPosition({
  project: "MyApp.Core",
  file: "UserService.cs",  // File name or path suffix (same matching as SearchSymbols)
  line: 42,                // 1-based
  column: 17               // Optional (1-based) — omit to resolve the most relevant symbol on the line
})

Returns: name, fullName, kind, signature, isDeclaration (whether the position sits on the symbol's own declaration), and (each omitted when empty/absent) containingType, documentation, definitionFile/Line (relative to resolvedPath's directory). Line-only queries prefer declarations on the line over referenced symbols.

Code Editing Tools (preview by default)

Surgical edits that emit a member-level change rather than a whole-file rewrite. Like ApplyFixes,
both default to preview mode (previewOnly: true) — nothing is written to disk unless you pass
previewOnly: false explicitly.

11. EditMember

Replace, add, or delete a single type member; returns a unified diff.

editMember({
  project: "MyApp.Core",
  symbol: "Acme.UserService.GetUser",  // The member (replace/delete), or the container type (add)
  operation: "replace",                 // "replace" | "add" | "delete"
  newSource: "public User GetUser(int id) => _repo.Find(id);",  // Required for replace/add
  previewOnly: false                    // Optional (default: true). Set false to write to disk.
})

Returns: operation, target, changedFiles, patch, previewOnly, applied, notes.

12. RenameSymbol

Rename a symbol and update every reference across the solution (Roslyn rename); returns a unified diff.

renameSymbol({ project: "MyApp.Core", symbol: "GetUser", newName: "GetUserById", previewOnly: false })

Returns: symbol, newName, changedFiles, patch, previewOnly, applied, verification, notes.

Every write is compile-verified. Before edit_member, rename_symbol or apply_fixes
touches disk, the candidate change is compiled in memory. If it introduces compiler errors —
including in a downstream project you never named — the write is refused: the call still
succeeds, applied comes back false, nothing is written, and you get the diff and the
errors. Pass allowIntroducedErrors: true to write anyway. A repository that was already
broken stays editable: the gate is what the change introduces, never whether the tree
compiles. See Compile Verification for the scope rule and
the guarantee boundary.

Tool Annotations

RoselineMCP's SDK (ModelContextProtocol 2.2.0) supports the standard MCP tool
annotation hints (readOnlyHint, destructiveHint,
idempotentHint), and every tool declares them via [McpServerTool(ReadOnly = ..., Destructive = ..., Idempotent = ...)]:

Tool	readOnlyHint	destructiveHint	idempotentHint	Notes
`AnalyzeSolution`	✅ true	❌ false	✅ true	Never writes to disk.
`ListDiagnostics`	✅ true	❌ false	✅ true	Never writes to disk.
`ApplyFixes`	❌ false	⚠️ true	❌ false	`destructiveHint` is a static, worst-case annotation: it's `true` because the tool can write files when `previewOnly: false` is passed, even though the default call (`previewOnly` unset, i.e. `true`) writes nothing. The SDK's annotation model has no way to express "destructive only for a specific parameter value" — see the doc comment on `ApplyFixesTool.ApplyFixes` in source.
`CheckCompilation`	✅ true	❌ false	✅ true	Never writes to disk. Compiles the loaded solution and reports the compiler's verdict.
`CreatePatch`	✅ true	❌ false	✅ true	Operates purely on the two provided strings; never touches the filesystem.
`SearchSymbols`	✅ true	❌ false	✅ true	Never writes to disk.
`GetSymbolInfo`	✅ true	❌ false	✅ true	Never writes to disk.
`FindReferences`	✅ true	❌ false	✅ true	Never writes to disk.
`FindImplementations`	✅ true	❌ false	✅ true	Never writes to disk.
`GetCallGraph`	✅ true	❌ false	✅ true	Never writes to disk.
`GetTypeHierarchy`	✅ true	❌ false	✅ true	Never writes to disk.
`GetSymbolAtPosition`	✅ true	❌ false	✅ true	Never writes to disk.
`EditMember`	❌ false	⚠️ true	❌ false	Same worst-case `destructiveHint` rationale as `ApplyFixes`: writes a file only when `previewOnly: false` is passed; the default call writes nothing.
`RenameSymbol`	❌ false	⚠️ true	❌ false	Same worst-case `destructiveHint` rationale as `ApplyFixes`: writes files only when `previewOnly: false` is passed; the default call writes nothing.

These hints are static per-tool metadata for MCP clients that surface them (e.g. to warn a user
before an agent invokes a destructive tool) — they describe the tool's worst-case behavior, not
the outcome of any specific call. See Tool Compatibility Policy
below for the stability guarantees around tool names and parameters that these annotations sit on
top of.

Tool Compatibility Policy

Tool names and required parameters are stable within a major version. An MCP client
integration written against AnalyzeSolution(pathOrGit, ...) on 1.x will keep working across
all 1.x releases.
Optional parameters may be added in minor versions (e.g. CreatePatch gained
ignoreWhitespace/ignoreCase as optional, defaulted parameters without breaking existing
callers).
Renaming or removing a parameter, changing a parameter's required/optional status, or
changing a tool's name is a breaking change. Breaking changes are called out under a
dedicated "Breaking Changes" heading in CHANGELOG.md and only ship in a major
version bump.
Response shapes (JSON field names/types) are documented in docs/API.md and
follow the same policy: additive fields are non-breaking, renamed/removed fields are breaking.

Supported Analyzers

The diagnostics tools (analyze_solution, list_diagnostics, apply_fixes) report compiler
diagnostics plus analyzer diagnostics, executed via Roslyn's CompilationWithAnalyzers
(check_compilation is deliberately not in this list — it is compiler-only, which is what makes
it fast enough for an edit loop):

Roslyn Analyzers -- Built-in C# compiler diagnostics
Roslynator -- 500+ analyzers and fixes for C#, bundled with RoselineMCP (shipped as an
analyzers/ folder next to RoselineMCP.dll) and executed by default, so RCS* diagnostics
surface and are fixable out of the box
Custom Analyzers -- Any Roslyn-based analyzer referenced by your analyzed solution is
loaded from the project's analyzer references and run alongside the bundled ones (there is no
built-in StyleCop.Analyzers reference — add it to your analyzed solution if you want SA*
diagnostics). Their code fixers are loaded too: an analyzer-reported rule is auto-fixable
when a matching fixer ships with Roslyn, in the bundled catalog, or inside one of the project's
own analyzer references (the bundled one wins for an ID both carry).

What could not be loaded is named. Roslyn reports an analyzer reference it cannot load — one
built against a newer Microsoft.CodeAnalysis than RoselineMCP's, a corrupt file — by
contributing zero analyzers, not by failing. The three diagnostics responses carry an
analyzerLoad block (analyzersRan, referencesConsulted, referencesContributing, analyzersLoaded, and a
notes[] entry per reference that contributed nothing: its name, the reason — load-failure
with Roslyn's errorCode and message, no C# analyzers, or exception). The block is omitted
when every reference contributed, so a present one always says something; see
docs/API.md.

Set RoselineMCP:RunAnalyzers to false for compiler-only diagnostics (faster on big
solutions; see Configuration and the analyzer-execution note in
SECURITY.md). The block is then present with analyzersRan: false, so "off"
stays distinguishable from "all fine".

Examples

Analyzing a Solution

# Using with an MCP client
mcp call analyzeSolution '{
  "pathOrGit": "/Users/dev/MyProject/MyProject.sln",
  "severity": "warning",
  "maxDiagnostics": 50
}'

Response:

{
  "solution": "MyProject.sln",
  "projects": 5,
  "diagnosticSummary": {
    "error": 2,
    "warning": 15,
    "info": 28
  },
  "topDiagnostics": [
    {
      "id": "CS0168",
      "severity": "warning",
      "message": "The variable 'ex' is declared but never used",
      "file": "Program.cs",
      "line": 42
    }
  ]
}

Applying Fixes

mcp call applyFixes '{
  "project": "MyProject.Core.csproj",
  "ids": ["CS0168", "RCS1001"],
  "previewOnly": true
}'

Response includes a unified diff patch showing all changes that would be applied.

Configuration

RoselineMCP reads appsettings.json and appsettings.{Environment}.json from the directory the
server binary is installed in (AppContext.BaseDirectory) — not from the process working
directory. Launching the server from inside a target repository never picks up that repository's
own appsettings.json, and the settings packaged with the dotnet tool install are always found.
Configuration is read once at startup; there is no reload-on-change file watching.

Environment Variables

Environment variables prefixed with ROSELINE_ override the JSON files, using __ as the section
separator. Settings under the RoselineMCP section therefore take a double prefix:

# RoselineMCP:EnableDiagnosticLogging
ROSELINE_RoselineMCP__EnableDiagnosticLogging=true

# RoselineMCP:DefaultTimeout (ms)
ROSELINE_RoselineMCP__DefaultTimeout=300000

# RoselineMCP:RunAnalyzers (compiler-only diagnostics when false)
ROSELINE_RoselineMCP__RunAnalyzers=false

# RoselineMCP:ConfirmDestructiveWrites (no elicitation before writes; unattended hosts)
ROSELINE_RoselineMCP__ConfirmDestructiveWrites=false

# RoselineMCP:ConfirmDestructiveWritesTimeout (ms to wait for that confirmation; 0 = forever)
ROSELINE_RoselineMCP__ConfirmDestructiveWritesTimeout=300000

# Logging:LogLevel:RoselineMCP
ROSELINE_Logging__LogLevel__RoselineMCP=Debug

DOTNET_ENVIRONMENT: Set environment (Development, Production)

appsettings.json

Configure logging and other settings:

{
  "Logging": {
    "LogLevel": {
      "Default": "Information",
      "RoselineMCP": "Debug"
    }
  },
  "RoselineMCP": {
    "DefaultTimeout": 120000,
    "EnableDiagnosticLogging": false,
    "WorkspaceCache": true,
    "RunAnalyzers": true,
    "ConfirmDestructiveWrites": true,
    "ConfirmDestructiveWritesTimeout": 300000
  }
}

RoselineMCP:DefaultTimeout: Wall-clock analysis budget (ms) applied to each tool call, in addition to the caller's own cancellation. 0 disables it. On the three write tools it does not cover the whole invocation: when a write confirmation is elicited, this clock starts only after the human answers, so think-time is charged to ConfirmDestructiveWritesTimeout instead. Size it for the analysis, not for the span a caller observes.
RoselineMCP:EnableDiagnosticLogging: Opt-in, local-only tracing of tool invocations — see Debug Logging. Disabled by default; enabled in appsettings.Development.json.
RoselineMCP:WorkspaceCache: Reuse the loaded MSBuild workspace across navigation/edit tool calls (enabled by default) — see Performance. Set to false to load a fresh workspace on every call. This is an isolation/debugging switch, not a memory-saving one — measured, disabling it costs ~26% more resident memory and ~45× second-call latency; see Memory Management.
RoselineMCP:RunAnalyzers: Run Roslyn analyzers (bundled Roslynator + the target project's own analyzer references) in the diagnostics tools (enabled by default) — see Supported Analyzers. Set to false for compiler-only diagnostics.
RoselineMCP:ConfirmDestructiveWrites: Ask the client to confirm (via MCP elicitation) before a write tool actually writes, when the caller passed previewOnly: false (enabled by default). Set to false for unattended hosts — CI, headless agents, or any client that cannot answer an elicitation — in which case the explicit previewOnly: false opt-in is the only guard before a write.
RoselineMCP:ConfirmDestructiveWritesTimeout: How long (ms) to wait for that confirmation before treating the silence as "no" and returning a preview (default 300000, 5 minutes). It is deliberately not DefaultTimeout — that is an analysis budget, and a human reading a real diff may legitimately exceed it. A client that accepts the prompt and never answers used to block the tool call forever; set this to 0 or less to restore that unbounded wait.

Architecture

RoselineMCP uses stdio transport — this is an intentional design decision. The server runs as a local process launched by the MCP client (Claude Desktop, AI agents), communicates over stdin/stdout, and exits when the client disconnects. This makes it perfectly suited for distribution as a NuGet global tool (dotnet tool install -g RoselineMCP) or Docker image — no port binding, no HTTP server, no infrastructure to manage.

┌─────────────────────┐
│   MCP Client        │
│  (Claude Desktop,   │
│   AI Assistants)    │
└────────┬────────────┘
         │ MCP Protocol (stdio)
         ▼
┌─────────────────────┐
│   RoselineMCP       │
│   MCP Server        │
│                     │
│  ┌───────────────┐  │
│  │  Tool Layer   │  │
│  │  (Analyze,    │  │
│  │   Fix, Patch) │  │
│  └───────┬───────┘  │
│          ▼          │
│  ┌───────────────┐  │
│  │ Service Layer │  │
│  │ (Workspace,   │  │
│  │  Diagnostics) │  │
│  └───────┬───────┘  │
│          ▼          │
│  ┌───────────────┐  │
│  │ Roslyn +      │  │
│  │ Roslynator    │  │
│  │ Analyzers     │  │
│  └───────────────┘  │
└─────────────────────┘
         │
         ▼
┌─────────────────────┐
│  C# Source Code     │
│  (.sln / .csproj)   │
└─────────────────────┘

Project Structure

RoselineMCP/
├── RoselineMCP/
│   ├── Interfaces/       # Service interfaces
│   ├── Services/         # Core business logic
│   ├── Tools/            # MCP tool implementations
│   ├── Models/           # Data transfer objects
│   └── Program.cs        # Application entry point
├── RoselineMCP.Tests/    # Unit tests
├── .github/workflows/    # CI/CD pipelines
├── Dockerfile            # Container build
└── RoselineMCP.sln       # Solution file

Performance

Workspace Cache -- The navigation, edit, and diagnostics/fix tools (ListDiagnostics,
ApplyFixes, and everything backed by IProjectLoader) reuse the loaded MSBuildWorkspace
across calls, cutting hundreds of milliseconds of reload off every call after the first. Cached
entries are fingerprinted (last-write-time + size of the .sln, every .csproj, and every
source file, plus their directories) and re-checked on each call, so any change on disk —
including RoselineMCP's own edits — triggers a fresh reload. A file/directory whose own last
write was still fresh when its stamp was captured also gets a content signature (a hash, or a
directory listing) recorded alongside it, so a same-length write landing in the same
filesystem-timestamp bucket is still caught by comparing that signature rather than trusted on a
bare stat match — cheaply, without forcing a reload on every call while the file stays fresh.
Disable with
RoselineMCP:WorkspaceCache = false — but note that this is an isolation/debugging switch, not
a way to reduce memory: a disposed workspace's memory is not returned to the OS, so disabling the
cache measures ~26% worse on resident memory as well as ~45× slower. The measured profile — and
why an idle-release was evaluated and rejected — is in
Memory Management
Workspace Isolation (AnalyzeSolution) -- AnalyzeSolution still creates a fresh
MSBuildWorkspace per operation (see Architecture)
Sequential Project Analysis -- Projects within a solution are analyzed one at a time, not
concurrently, to keep MSBuild workspace state consistent
Result Capping -- maxDiagnostics/max bound how many diagnostics are returned per call,
independent of how many were found

Compile guard

The compile gate behind EditMember/RenameSymbol/ApplyFixes only fires on writes made through
those tools. The compile guard puts the same verdict behind every file write in a C# solution —
Edit, Write, MultiEdit, whatever your agent reached for — by running as a PostToolUse hook
that asks the already-running server for a verdict over a local socket.

It reports; it cannot block. PostToolUse fires after the tool has written, and the hook
contract carries no blocking decision. What the guard does is surface the introduced compiler errors
to the agent inside the same turn, so the next thing it does is repair them instead of building on
top of them. Preventing the write is not on offer — see
the issue for why the alternative
that can block was rejected.

Turn it on (it is off by default), then install the hook:

# 1. the server opens the local guard endpoint
export ROSELINE_RoselineMCP__Guard=true

# 2. print the settings.json block and paste it into .claude/settings.json
roseline-mcp guard --print-hook

What it does on each write:

Situation	Result
The edit introduced compiler errors	the errors, in the agent's turn (exit `2`, stderr)
The edit introduced nothing	silence — exit `0`, no output, no tokens
The branch was already red	silence — pre-existing errors are not the agent's doing
Not a `.cs` file, or no owning project	silence
Server not running, or slow to answer	silence — a guard that cannot inform must not interrupt

Compiler diagnostics only, and scoped to the changed project plus its dependents, so it is fast
enough to sit in an inner loop. It adds a local IPC endpoint: read
SECURITY.md before enabling it on a shared
machine.

Security

Read-Only by Default -- AnalyzeSolution, ListDiagnostics, CreatePatch, and all six code
navigation tools (SearchSymbols, GetSymbolInfo, FindReferences, FindImplementations,
GetCallGraph, GetTypeHierarchy) never write to disk. The three write-capable tools —
ApplyFixes, EditMember, and RenameSymbol — each default to previewOnly: true; writing
requires the caller to pass previewOnly: false explicitly. Behind that opt-in sits a second,
best-effort guard: the write tools ask the client to confirm via MCP elicitation before writing,
and a decline downgrades the call to a preview. So does an unanswered prompt — the round-trip is
bounded by RoselineMCP:ConfirmDestructiveWritesTimeout (default 5 minutes), after which the call
returns a preview rather than writing or hanging. That confirmation is skipped when the client
cannot elicit, and an operator can disable it outright with
RoselineMCP:ConfirmDestructiveWrites=false — after which previewOnly: false is the only guard
left. See SECURITY.md.
Real, Read-Only Git Cloning -- pathOrGit accepts http(s):// Git URLs, which are
shallow-cloned (git clone --depth 1) into a temp directory that's deleted after the operation.
No other URL scheme is treated as a Git remote.
MSBuild Is Not a Sandbox -- loading a .sln/.csproj via MSBuildWorkspace is a
design-time MSBuild evaluation and can execute build logic embedded in the project (<Exec>
tasks, custom UsingTask assemblies, imported .targets/.props). Analyzing a fully untrusted
repository or URL carries a real code-execution risk on the host. See
SECURITY.md for the full write-up and operator recommendations before
pointing RoselineMCP at untrusted input.
Analyzer Execution Is Code Execution -- the diagnostics tools also run the target project's
own referenced Roslyn analyzers in-process (see
Supported Analyzers); an analyzer from an untrusted repository is
arbitrary code. RoselineMCP:RunAnalyzers = false disables the diagnostic analyzer pass
(bundled Roslynator included). See SECURITY.md.
Code-Fix Providers Come From Those Same References -- list_diagnostics and apply_fixes
also load the CodeFixProvider types a project's own analyzer references carry. That adds no
assembly the analyzer pass does not already load (each is obtained through the reference's own
assembly loader), but it is a decision rather than an accident, and RunAnalyzers = false does
not govern it. Recorded in SECURITY.md.
Source Generators Run Regardless Of That Switch -- generators ship through the same
AnalyzerReferences, but run as part of building any compilation rather than as part of the
diagnostics pass, so every semantic path executes them -- all navigation tools included. They
cannot be suppressed without breaking semantic analysis. RunAnalyzers = false narrows the
code-execution surface of an untrusted repository; it does not close it. Isolate instead --
see SECURITY.md.
No Dedicated Path-Traversal Sandbox -- paths are resolved with plain existence checks, not
canonicalized against an allowed root; treat pathOrGit/project/branch as trusted operator
input.

Troubleshooting

Common Issues

MSBuild not found: Ensure .NET SDK is installed and in PATH
Solution won't load: Check for missing NuGet packages, run dotnet restore
No analyzer (RCS/SA) diagnostics found**: Verify RoselineMCP:RunAnalyzers isn't set to false; for non-bundled rule sets (e.g. StyleCop), verify the analyzer is installed in the target project
Permission denied: Ensure read access to solution files

Debug Logging

Enable detailed logging:

ROSELINE_Logging__LogLevel__RoselineMCP=Debug dotnet run --project RoselineMCP/RoselineMCP.csproj

Tracing Individual Tool Calls

Every tool call gets a per-invocation correlation ID (a GUID). It's cheap to generate so it's
always created, but it's only surfaced when you need it: it's included in every JSON error
response (correlationId) and attached to that call's log lines via ILogger.BeginScope, so a
user reporting a failure can hand you one ID that ties back to the full server-side log entry —
without needing to grep timestamps.

For deeper, opt-in tracing of each tool invocation (start/stop, duration, success/failure) as a
System.Diagnostics.Activity span, set RoselineMCP:EnableDiagnosticLogging to true (it's
already true in appsettings.Development.json):

ROSELINE_RoselineMCP__EnableDiagnosticLogging=true dotnet run --project RoselineMCP/RoselineMCP.csproj

This uses the built-in ActivitySource/Activity APIs rather than the OpenTelemetry SDK, so it
adds no extra dependency. Spans are logged exclusively through the existing ILogger pipeline,
which is already routed to stderr — never to stdout (the MCP JSON-RPC channel) — and nothing is
ever sent over the network; when the flag is off (the default), no listener is registered and the
spans cost essentially nothing.

Roadmap

Additional analyzer rule sets (SonarAnalyzer, FxCop)
Auto-fix suggestions with confidence scoring
CI/CD integration for automated analysis pipelines
Multi-solution support in a single session
Incremental analysis (only changed files)
Custom analyzer rule configuration via MCP

Want to contribute? Pick any roadmap item and open a PR!

Documentation

Documentation site -- overview, tool
reference, and the token-savings benchmark
(built from website/, deployed to GitHub Pages)
docs/API.md -- Full request/response reference for every MCP tool, service
interfaces, models, and the error-response contract
docs/AGENT-BENCHMARK.md -- End-to-end A/B: does an AI agent actually
spend fewer tokens with RoselineMCP? (honest answer — a large-codebase win, break-even on small)
docs/ARCHITECTURE.md -- Layered architecture, data flow, and design
patterns
PROMPTS.md -- Example prompts and end-to-end workflows for each tool
CHANGELOG.md -- Release history and breaking changes
SECURITY.md -- Vulnerability reporting and the MSBuild code-execution caveat
CONTRIBUTING.md -- Development setup and PR process

Atypical MCP servers

Part of a suite of Model Context Protocol servers by Atypical Consulting:

RoselineMCP — Roslyn code intelligence for AI agents
ASTral — structured code retrieval (tree-sitter)
AdrMcp — Architecture Decision Records over MCP
MarkdownInk — Markdown rendering for the terminal

Contributing

Contributions are welcome! Please read CONTRIBUTING.md first.

Fork the repository
Create your feature branch (git checkout -b feature/amazing-feature)
Commit using conventional commits (git commit -m 'feat: add amazing feature')
Push to the branch (git push origin feature/amazing-feature)
Open a Pull Request

License

Acknowledgments

Built on Roslyn -- The .NET Compiler Platform
Powered by Roslynator -- C# analyzers and refactorings
Uses DiffPlex -- Diff generation library
Implements Model Context Protocol -- AI assistant integration protocol

Built with care by Atypical Consulting -- opinionated, production-grade open source.

Built by Atypical Consulting. We also make
NuGetKeep,
a self-hosted NuGet server with supply-chain quarantine.