MCP server that connects AI assistants to Visual Studio's Roslyn compiler for 100% accurate C# code analysis.
Platform: Windows only | Language: C# only
Tired of typing prompts manually? 🎤
Try Murmur — offline voice-to-text for Windows. Fast, private, no cloud.
👉 https://murmurvt.com/
☕ Like it? Buy me a coffee so I can mass-produce more extensions at 3 AM → ko-fi.com/yaroslavhorokhov
⚡ TL;DR: How to Connect
- Open your C# solution in Visual Studio
- Go to View → Other Windows → RoslynMCP
- Copy the JSON configuration to your AI assistant's config
Example for Claude Code:
claude mcp add roslyn -- "c:\users\<username>\appdata\local\microsoft\visualstudio\<vs-instance>\extensions\<extension-id>\Proxy\RoslynMcp.Proxy.exe"
⚠️ The path is unique to your VS installation — don't copy the example above, copy from the RoslynMCP tool window.
Multiple VS instances: Each Visual Studio window runs its own MCP server. Your AI assistant automatically connects to the one with your solution's folder open.
Database: Creates .roslyn-mcp/memory.db (SQLite) in your solution directory for memory, knowledge base, and graph data.
First time? Copy this and send to your AI assistant:
Save this rule permanently using memory_learn with scope:"global":
RoslynMCP basics:
1. Find & Execute Tools: 80+ tools available
- search_tools("what you need") → find the right tool
- call_tool(name, args) → execute it
2. C# Code: Never grep/Glob for .cs files — only Roslyn tools (100% semantic accuracy)
3. Memory System (3 layers):
- memory_* — context & rules (scope:"session" for temp, scope:"global" for permanent)
- kb_* — knowledge base (ideas, docs, history, links — lives forever)
- graph_* — track changes, dependencies, cause→effect relationships
4. Start conversation with: memory_context to recall saved rules + context
Command:
memory_learn(content:"RoslynMCP: search_tools to find, call_tool to execute (80+ tools). Never grep for .cs - Roslyn only. Memory layers: memory_* (session), kb_* (permanent knowledge), graph_* (dependencies/changes). Start with memory_context.", scope:"global", importance:10, key:"roslyn-basics")
This saves rules permanently. Next conversation, AI will auto-load them via memory_context.
Requirements
- Windows 10/11
- Visual Studio 2022 (17.9 or later)
- C# solution (.sln)
Supported AI Assistants
- Claude Code (Anthropic)
- Codex
- Any MCP-compatible client
Why RoslynMCP?
AI assistants use text search (grep/ripgrep) which misses semantic relationships. RoslynMCP provides compiler-level accuracy using the same Roslyn engine that powers Visual Studio IntelliSense.
| Capability |
RoslynMCP |
Text Search |
| Symbol resolution |
Semantic (100%) |
Text matching (name collisions) |
| Overloads |
Distinguishes all |
Cannot distinguish |
| Type hierarchy |
Full inheritance tree |
Not available |
| Call graph |
Callers & callees |
Not available |
| Refactoring |
Safe rename across solution |
Find & replace (risky) |
Use RoslynMCP: symbol usages, type hierarchies, call graphs, safe refactoring, impact analysis, persistent memory
Installation
- Install RoslynMCP from Visual Studio Marketplace
- Open your C# solution in Visual Studio
- Go to View → Other Windows → RoslynMCP
- Copy the configuration for your AI assistant
Features
v1.17.8 — Memory Scope & Retrieval v2
- Updated: Session-first memory policy —
memory_remember and memory_learn default to session scope.
- Improved: Global scope guardrails with explicit scope adjustment reason for non-reusable entries.
- Enhanced:
memory_context retrieval v2 with scoped candidates, hybrid vector + FTS search, and rerank metadata.
- New: Scope lifecycle tools —
memory_promote, memory_audit_scopes, memory_rebalance_scopes.
- Fixed: Stable memory search quality via FTS tag sync and embedding refresh on content/tag updates.
- Added: Clear
text_search usage guidance for LLMs (queryMode, wholeWord, filePattern, maxResults, regexTimeoutMs).
- Fixed: TextSearch workflow docs now clarify safe defaults (
queryMode:"literal" first), truncation handling, and wildcard pitfalls (*term* over broad scope).
- Fixed: File-path handling in Roslyn file tools (
reload_file, get_file_outline, get_types_in_file, preview_extract_method) now resolves relative/absolute paths consistently.
- Fixed:
apply_move_type now applies workspace changes on UI thread (prevents TryApplyChanges background-thread failure).
- Fixed:
apply_split_class now supports overloaded methods without duplicate-key failures.
- Fixed:
organize_usings rewritten to deterministic safe top-level using sort/remove flow (no destructive import pass).
- Fixed:
memory_search FTS snippets now return content snippets (not memory IDs).
- Fixed:
memory_end_session duration uses robust UTC round-trip parsing and clamps negative durations.
New: memory_context now accepts sessionId parameter to query specific session memories.
Fixed: Removed misleading default values from memory tool schemas — AI can freely choose which parameters to pass.
AI Prompts for Memory
Global rule (forever):
Remember forever: [rule]
Save as global rule: [rule]
Current session:
Note for this session: [note]
Remember for now: [decision]
Knowledge base (permanent):
Save to KB: [info]
Add to knowledge base as idea/note/link: [content]
Sessions:
Start session: [goal]
End session: [summary]
Recall:
What do you remember about [topic]?
Show KB history
What was in the last session?
v1.17.6 — JSON-RPC Case-Insensitive Parsing
Fixed: JSON-RPC deserialization now accepts both camelCase and PascalCase property names (e.g., both "jsonrpc" and "JsonRpc"). This improves compatibility with various MCP clients.
v1.17.5 — Documentation & Onboarding
Improved: Added "TL;DR: How to Connect" section with quick setup guide, AI teaching instructions, and memory_learn command for permanent rules.
Fixed: Global memories (scope:"global") no longer get sessionId attached — they are truly session-independent now.
v1.17.4 — Graph Build Counters Fix
Fixed: graph_build_from_type now correctly counts nodes and edges. nodesSkipped renamed to edgesSkipped.
Fixed: search_tools now uses custom descriptions from config_tool_enabled.
v1.17.2 — Knowledge Base (Persistent Long-Term Memory)
New kb_* tools for storing knowledge that persists across sessions forever.
Memory vs Knowledge Base
| Feature |
memory_* (Session Memory) |
kb_* (Knowledge Base) |
| Purpose |
Current work context |
Long-term knowledge |
| Lifetime |
Session or until deleted |
Forever (persistent) |
| Use for |
Task context, temp notes, decisions during work |
History, ideas, docs, links, snippets, todos |
| Search |
By type, tags, files |
By type, category, tags, dates + semantic |
Use memory_* for: what you're doing now, session context, temporary decisions
Use kb_* for: what you want to remember forever — history, ideas, documentation, useful links
Knowledge Base Types
| Type |
Use for |
Note |
Observations, thoughts, meeting notes |
Idea |
Future features, "we could...", backlog |
Doc |
How-to guides, setup instructions |
Snippet |
Reusable code, templates, patterns |
Link |
URLs with context (docs, articles) |
History |
"We did X", milestones, changelog |
Todo |
Non-urgent tasks, "don't forget to..." |
Reference |
API docs, config explanations |
| Tool |
Purpose |
kb_add |
Create entry (type, category, title, content, url?, tags?) |
kb_get |
Get entry by ID (with full content) |
kb_update |
Update any field individually (only passed fields change) |
kb_delete |
Delete entry (cascade? for children) |
kb_search |
Semantic + FTS search with filters (type, category, dates) |
kb_list |
List entries with filters (returns summaries by default) |
kb_tree |
View hierarchy (parent/children) |
kb_related |
Find related entries |
Token-Efficient Workflow
# Get summaries first (no content = less tokens)
kb_list type:"History" createdAfter:"2025-12-22"
# Then fetch full content for specific entries
kb_get id:"abc123"
Date filters:
"what did we do yesterday" → createdAfter:"2025-12-22" createdBefore:"2025-12-23""last week" → createdAfter:"2025-12-16"
- Fixed: AlwaysLoadedTools now show full JSON schema (was minimal
{type:object})
- Optimized: Removed
validate_text, reload_file, get_errors from always-loaded (find via search_tools)
- New:
RegenerateToolVector — vectors update immediately when description changes
v1.17.0 — Teaching Claude
- Fixed:
memory_forget_session(deleteSession=true) now correctly deletes ALL session memories regardless of keepGlobal flag
- New: Added guidance for teaching Claude how to use RoslynMCP effectively via memory
- Architecture: Only 4.6k tokens for 86 tools via
search_tools → call_tool pattern (vs ~40k+ if all tools loaded)
All 86 tool descriptions are indexed as 384-dimensional vectors using all-MiniLM-L6-v2 ONNX model. When you call search_tools("find usages"), the query is embedded and matched via cosine similarity.
search_tools("find usages") → embedding → cosine similarity → top-K tools
Config options (via config_set):
| Key | Default | Description |
|-----|---------|-------------|
| tool_search_threshold | 0.3 | Minimum similarity score (0-1) |
Two Graph Systems
RoslynMCP uses two complementary graph systems:
| Graph |
Tools |
Purpose |
| Roslyn Call Graph |
find_callers, find_callees, understand_method |
Static code analysis — who calls what |
| DevGraph |
graph_* tools |
Dynamic tracking — changes, causes, decisions |
Roslyn Graph is built automatically from code. Use it to navigate:
find_callers("SaveUser") → which methods call SaveUser?
find_callees("SaveUser") → what does SaveUser call?
understand_method("SaveUser") → callers + callees + body in one call
DevGraph is built during work. Use it to track:
graph_track_change("User.cs", "Added validation")
graph_track_cause(errorNodeId, "Fixed null reference")
graph_get_impact("UserService") → what breaks if I change this?
Teaching Claude via Memory
Write rules to memory so Claude remembers how to work with your codebase. Use memory_learn for global rules:
memory_learn("For C# code: ONLY use Roslyn MCP tools, NEVER grep/Glob. validate_text BEFORE edit, reload_file AFTER.")
memory_learn("Write memories in English (better vector search). Show user info in their language.")
memory_learn("Use graph tools with sessions: start_session before tracking, graph_track_change after edits, graph_track_cause for bug fixes.")
Recommended starter memories:
| Rule | Why |
|------|-----|
| Use Roslyn tools for C# | 100% accurate vs text search |
| validate_text before edit | Catch errors before saving |
| reload_file after edit | Sync Roslyn workspace |
| English in memory | Vector search works better |
| Graph with sessions | Track changes, dependencies, causes |
After memory_learn, Claude will recall these rules via memory_context at conversation start.
v1.16.2
- New:
config_tool_enabled — Enable/disable tools to save memory. Disabled tools won't load after server restart
- New: Custom tool descriptions — LLM can update tool descriptions stored in DB (used in
tools/list response)
- New:
config_list now shows all tools with enabled/disabled status and stats
v1.16.1
- Improved:
get_dependency_graph — Full project graph with edges by default. New params: rootProject (filter), maxDepth, direction (dependencies/dependents/both) for large solutions (70+ projects)
- Fixed:
memory_* tools now accept string importance ("high"→8, "medium"→5, "low"→3) in addition to numbers
- Fixed:
memory_update accepts both id and memoryId parameters
v1.16.0 — Memory & Graph (Experimental)
- New:
memory_* tools (14 tools) — Persistent memory across sessions
- New:
graph_* tools (10 tools) — Track dependencies, changes, cause-effect relationships
- New:
graph_build_from_type — Auto-build dependency graph from Roslyn analysis (hierarchy, callers, callees)
- Database:
.roslyn-mcp/memory.db (SQLite) in your solution directory
- Vector Search: Uses ONNX model (
all-MiniLM-L6-v2) for semantic memory search
- Config: LLM can tune vector search via
config_set (e.g., vector_threshold, scoring_similarity_weight)
⚠️ Experimental: Memory and Graph tools are new. We're still figuring out the best workflows. Try different approaches and find what works for you!
Memory Scopes & Importance
| Scope |
Lifetime |
Use for |
global |
Forever |
Project rules, patterns, architectural decisions |
session |
Until session ends |
Current task context, temporary notes |
Importance (1-10): Higher = more likely to appear in memory_context. Use 8-10 for critical rules, 5-7 for useful info, 1-4 for minor notes.
| Tool |
Purpose |
memory_start_session |
Start tracking a dev session |
memory_end_session |
End session with summary |
memory_list_sessions |
List all sessions |
memory_restore |
Restore context from previous session |
memory_remember |
Save decision/change/error/fix |
memory_learn |
Save pattern/lesson (global) |
memory_recall |
Get memories by type/tags/files |
memory_context |
Get relevant context for task |
memory_search |
Semantic search in memories |
memory_update |
Update existing memory |
memory_consolidate |
Merge similar memories |
memory_analyze |
Memory health stats |
memory_forget |
Delete a memory |
memory_forget_session |
Delete session + its graph nodes |
| Tool |
Purpose |
graph_track_change |
Track file modification |
graph_track_dependency |
Track A depends on B |
graph_track_cause |
Track cause → effect |
graph_build_from_type |
Auto-build graph from Roslyn (hierarchy + calls) |
graph_get_impact |
What breaks if I change X? |
graph_get_dependencies |
Show dependency tree |
graph_get_path |
Find path between two nodes |
graph_get_history |
File change history |
graph_visualize |
Generate Mermaid diagram |
graph_delete_node |
Remove node + orphans |
Quick Start
# Start session
memory_start_session("Adding feature X")
# Track changes as you work
graph_track_change("MyClass.cs", "Added new method")
graph_track_dependency("ServiceA", "ServiceB")
# Or auto-build from existing code
graph_build_from_type("MyService") # builds hierarchy + callers + callees
# Save learnings
memory_learn("Always validate input before processing")
# End session
memory_end_session("Feature X completed")
# Later: recall context
memory_context(forTask="similar feature")
graph_get_impact(nodeId="ServiceB")
v1.14.1
- Optimized: Reduced AI context usage (from ~40K to ~33K tokens) by streamlining tool descriptions
- Removed:
change_scope (use impact_analysis), find_circular_dependencies, list_refactoring_services - rarely used tools
v1.13.2
- New:
extract_interface - Extract interface from class for DI/testability. Generates IClassName.cs and adds interface implementation to class
- New:
organize_usings - Sort and remove unused using directives. Roslyn-powered cleanup with CS8019/CS0105 diagnostics
v1.13.0
- New:
preview_split_class / apply_split_class - Split large classes into partial files with smart method grouping
- New:
preview_move_type / apply_move_type - Move types to their own files
- New:
apply_extract_method - Extract code selection into a new method
- Improved:
get_type_members - Now works with ANY type including external (NuGet, System., Microsoft.). Use short name (Solution) or full name (Microsoft.CodeAnalysis.Solution)
- Improved: Tool descriptions with concrete examples for better AI tool selection
- Removed:
get_completions, get_signature_help - Replaced by enhanced get_type_members and get_overloads
v1.12.0
- New:
get_full_context - Get complete context for a symbol in one call: callers, callees, dependencies, related types
- Improved: Redesigned all tool descriptions for better AI tool selection - priorities, explicit negations, use cases, workflow guidance
v1.11.0
- New:
get_completions - Get code completion suggestions at any position (like IntelliSense)
- New:
get_signature_help - Get method signature and parameter info during calls
- New:
apply_rename - Rename symbols across entire solution with Roslyn accuracy
v1.10.0
- Fixed:
reload_file now works correctly (was failing with "TryApplyChanges cannot be called from a background thread")
RoslynMCP provides 90+ Roslyn-powered tools:
| Category |
Tools |
| 🔍 Navigation |
find_references, find_definition, find_callers, find_callees, find_implementations, find_overrides, find_base_members |
| 🧠 Understanding |
understand_type, understand_method, get_type_info, get_type_members, get_method_body, get_class_hierarchy |
| 📁 Structure |
get_solution_structure, get_project_structure, get_file_outline, get_types_in_file, find_entry_points, get_dependency_graph |
| 🩺 Diagnostics |
get_errors, get_warnings, validate_text, find_async_issues, find_performance_issues, find_unused_code |
| 🔧 Refactoring |
preview_rename, apply_rename, extract_interface, organize_usings, impact_analysis, preview_extract_method, suggest_refactorings |
| 💡 IntelliSense |
get_quick_fixes, get_overloads, get_xml_documentation |
| 🔎 Search |
find_attribute_usages, find_event_subscribers, find_extension_methods, find_tests_for_type, text_search |
| 📊 Metrics |
get_code_metrics, analyze_data_flow, get_full_context, get_constructor_parameters |
| 🧠 Memory |
memory_start_session, memory_end_session, memory_remember, memory_learn, memory_recall, memory_context, memory_search, memory_forget, memory_forget_session, memory_update, memory_consolidate, memory_analyze, memory_list_sessions, memory_restore |
| 📚 Knowledge Base |
kb_add, kb_get, kb_update, kb_delete, kb_search, kb_list, kb_tree, kb_related |
| 🔗 Graph |
graph_track_change, graph_track_dependency, graph_track_cause, graph_build_from_type, graph_get_impact, graph_get_history, graph_get_dependencies, graph_get_path, graph_visualize, graph_delete_node |
Troubleshooting
Extension not loading
- Check Visual Studio Output window for errors
- Verify extension is enabled in Extensions → Manage Extensions
AI assistant not connecting
- Ensure Visual Studio has a solution open
- Open the RoslynMCP tool window (View → Other Windows → RoslynMCP)
- Verify the server status shows "Running"
- Make sure you copied the correct configuration for your AI assistant
- Restart your AI assistant after configuration changes
License
PolyForm Noncommercial 1.0.0 — Free for personal and non-commercial use.
Commercial use requires a separate license. See LICENSE for details.
For commercial licensing, contact: ko-fi.com/yaroslavhorokhov
