Code Graph Context

Give your AI coding assistant a photographic memory of your codebase.

Code Graph Context is an MCP server that builds a semantic graph of your TypeScript codebase, enabling Claude to understand not just individual files, but how your entire system fits together.

Config-Driven & Extensible: Define custom framework schemas to capture domain-specific patterns beyond the included NestJS support. The parser is fully configurable to recognize your architectural patterns, decorators, and relationships.

                    ┌─────────────────────────────────────────────────────────────┐
                    │                     YOUR CODEBASE                           │
                    │  ┌──────────┐  ┌──────────┐  ┌──────────┐  ┌──────────┐    │
                    │  │ Service  │  │Controller│  │  Module  │  │  Entity  │    │
                    │  └────┬─────┘  └────┬─────┘  └────┬─────┘  └────┬─────┘    │
                    └───────┼─────────────┼─────────────┼─────────────┼──────────┘
                            │             │             │             │
                            ▼             ▼             ▼             ▼
                    ┌─────────────────────────────────────────────────────────────┐
                    │                   CODE GRAPH CONTEXT                        │
                    │                                                             │
                    │   AST Parser ──► Neo4j Graph ──► Vector Embeddings          │
                    │   (ts-morph)     (Relationships)  (Local or OpenAI)         │
                    │                                                             │
                    └─────────────────────────────────────────────────────────────┘
                                                │
                                                ▼
                    ┌─────────────────────────────────────────────────────────────┐
                    │                      CLAUDE CODE                            │
                    │                                                             │
                    │   "What services depend on UserService?"                    │
                    │   "What's the blast radius if I change this function?"      │
                    │   "Find all HTTP endpoints that accept a UserDTO"           │
                    │   "Refactor this across all 47 files that use it"           │
                    │                                                             │
                    └─────────────────────────────────────────────────────────────┘

Why Code Graph Context?

Without Code Graph	With Code Graph
Claude reads files one at a time	Claude understands the entire dependency tree
"What uses this?" requires manual searching	Instant impact analysis with risk scoring
Refactoring misses edge cases	Graph traversal finds every reference
Large codebases overwhelm context	Semantic search finds exactly what's relevant
Multi-file changes are error-prone	Swarm agents coordinate parallel changes

Features

• Multi-Project Support: Parse and query multiple projects in a single database with complete isolation
• Semantic Search: Vector-based search using local or OpenAI embeddings to find relevant code
• Natural Language Querying: Convert questions into Cypher queries
• Framework-Aware: Built-in NestJS schema with ability to define custom framework patterns
• Weighted Graph Traversal: Intelligent traversal scoring paths by importance and relevance
• Workspace Support: Auto-detects Nx, Turborepo, pnpm, Yarn, and npm workspaces
• Parallel & Async Parsing: Multi-threaded parsing with Worker threads for large codebases
• Streaming Import: Chunked processing for projects with 100+ files
• Incremental Parsing: Only reparse changed files
• File Watching: Real-time graph updates on file changes
• Impact Analysis: Assess refactoring risk (LOW/MEDIUM/HIGH/CRITICAL)
• Dead Code Detection: Find unreferenced exports with confidence scoring
• Duplicate Detection: Structural (AST hash) and semantic (embedding similarity) duplicates
• Swarm Coordination: Multi-agent stigmergic coordination with pheromone decay

Architecture

TypeScript Source → AST Parser (ts-morph) → Neo4j Graph + Vector Embeddings → MCP Tools

Core Components:

• src/core/parsers/typescript-parser.ts - AST parsing with ts-morph
• src/storage/neo4j/neo4j.service.ts - Graph storage and queries
• src/core/embeddings/embeddings.service.ts - Embedding service (local sidecar or OpenAI)
• src/mcp/mcp.server.ts - MCP server and tool registration

Dual-Schema System:

• Core Schema: AST-level nodes (ClassDeclaration, MethodDeclaration, ImportDeclaration, etc.)
• Framework Schema: Semantic interpretation (NestController, NestService, HttpEndpoint, etc.)

Nodes have both coreType (AST) and semanticType (framework meaning), enabling queries like "find all controllers" while maintaining AST precision.

Quick Start

Prerequisites

• Node.js >= 18
• Python >= 3.10 (for local embeddings)
• Docker (for Neo4j)

No API keys required. Local embeddings work out of the box using a Python sidecar.

1. Install

npm install -g code-graph-context
code-graph-context init  # Sets up Neo4j + Python sidecar + downloads embedding model

The init command handles everything:

• Starts a Neo4j container via Docker
• Creates a Python virtual environment
• Installs embedding dependencies (PyTorch, sentence-transformers)
• Downloads the default embedding model (~3GB)

2. Configure Claude Code

claude mcp add --scope user code-graph-context -- code-graph-context

That's it. No API keys needed. Restart Claude Code and you're ready to go.

Want to use OpenAI instead? See Embedding Configuration below.

3. Parse Your Project

In Claude Code, say:

"Parse this project and build the code graph"

Claude will run parse_typescript_project and index your codebase.

---

Configuration Files

Claude Code stores MCP server configs in JSON files. The location depends on scope:

Scope	File	Use Case
User (global)	~/.claude.json	Available in all projects
Project	.claude.json in project root	Project-specific config
Local	.mcp.json in project root	Git-ignored local overrides

Manual Configuration

If you prefer to edit the config files directly:

~/.claude.json (user scope - recommended):

{
  "mcpServers": {
    "code-graph-context": {
      "command": "code-graph-context"
    }
  }
}

With OpenAI (optional):

{
  "mcpServers": {
    "code-graph-context": {
      "command": "code-graph-context",
      "env": {
        "OPENAI_EMBEDDINGS_ENABLED": "true",
        "OPENAI_API_KEY": "sk-your-key-here"
      }
    }
  }
}

From source installation:

{
  "mcpServers": {
    "code-graph-context": {
      "command": "node",
      "args": ["/absolute/path/to/code-graph-context/dist/cli/cli.js"]
    }
  }
}

Environment Variables

Variable	Required	Default	Description
NEO4J_URI	No	bolt://localhost:7687	Neo4j connection URI
NEO4J_USER	No	neo4j	Neo4j username
NEO4J_PASSWORD	No	PASSWORD	Neo4j password
EMBEDDING_MODEL	No	codesage/codesage-base-v2	Local embedding model (see Embedding Configuration)
EMBEDDING_BATCH_SIZE	No	8	Texts per embedding batch (lower = less memory, higher = faster)
EMBEDDING_SIDECAR_PORT	No	8787	Port for local embedding server
EMBEDDING_DEVICE	No	auto (mps/cpu)	Device for embeddings. Auto-detects MPS on Apple Silicon
EMBEDDING_HALF_PRECISION	No	false	Set true for float16 (uses ~0.5x memory)
OPENAI_EMBEDDINGS_ENABLED	No	false	Set true to use OpenAI instead of local embeddings
OPENAI_API_KEY	No*	-	Required when OPENAI_EMBEDDINGS_ENABLED=true; also enables natural_language_to_cypher

---

Core Capabilities

Semantic Code Search

Find code by describing what you need, not by memorizing file paths:

"Find where user authentication tokens are validated"
"Show me the database connection pooling logic"
"What handles webhook signature verification?"

Impact Analysis

Before you refactor, understand the blast radius:

┌─────────────────────────────────────────────────────────────┐
│ Impact Analysis: UserService.findById()                     │
├─────────────────────────────────────────────────────────────┤
│ Risk Level: HIGH                                            │
│                                                             │
│ Direct Dependents (12):                                     │
│   └── AuthController.login()                                │
│   └── ProfileController.getProfile()                        │
│   └── AdminService.getUserDetails()                         │
│   └── ... 9 more                                            │
│                                                             │
│ Transitive Dependents (34):                                 │
│   └── 8 controllers, 15 services, 11 tests                  │
│                                                             │
│ Affected Files: 23                                          │
│ Recommendation: Add deprecation warning before changing     │
└─────────────────────────────────────────────────────────────┘

Graph Traversal

Explore relationships in any direction:

UserController
    │
    ├── INJECTS ──► UserService
    │                   │
    │                   ├── INJECTS ──► UserRepository
    │                   │                   │
    │                   │                   └── MANAGES ──► User (Entity)
    │                   │
    │                   └── INJECTS ──► CacheService
    │
    └── EXPOSES ──► POST /users
                        │
                        └── ACCEPTS ──► CreateUserDTO

Dead Code Detection

Find code that can be safely removed:

Dead Code Analysis: 47 items found
├── HIGH confidence (23): Exported but never imported
│   └── formatLegacyDate() in src/utils/date.ts:45
│   └── UserV1DTO in src/dto/legacy/user.dto.ts:12
│   └── ... 21 more
├── MEDIUM confidence (18): Private, never called
└── LOW confidence (6): May be used dynamically

Duplicate Code Detection

Identify DRY violations across your codebase:

Duplicate Groups Found: 8

Group 1 (Structural - 100% identical):
├── validateEmail() in src/auth/validation.ts:23
└── validateEmail() in src/user/validation.ts:45
    Recommendation: Extract to shared utils

Group 2 (Semantic - 94% similar):
├── parseUserInput() in src/api/parser.ts:78
└── sanitizeInput() in src/webhook/parser.ts:34
    Recommendation: Review for consolidation

---

Swarm Coordination

Execute complex, multi-file changes with parallel AI agents.

The swarm system enables multiple Claude agents to work on your codebase simultaneously, coordinating through the code graph without stepping on each other.

                         ┌──────────────────┐
                         │   ORCHESTRATOR   │
                         │                  │
                         │ "Add JSDoc to    │
                         │  all services"   │
                         └────────┬─────────┘
                                  │
                    ┌─────────────┼─────────────┐
                    │             │             │
                    ▼             ▼             ▼
             ┌──────────┐  ┌──────────┐  ┌──────────┐
             │ Worker 1 │  │ Worker 2 │  │ Worker 3 │
             │          │  │          │  │          │
             │ Claiming │  │ Working  │  │ Claiming │
             │ AuthSvc  │  │ UserSvc  │  │ PaySvc   │
             └──────────┘  └──────────┘  └──────────┘
                    │             │             │
                    └─────────────┼─────────────┘
                                  │
                                  ▼
                    ┌─────────────────────────────┐
                    │      PHEROMONE TRAILS       │
                    │                             │
                    │  AuthService: [claimed]     │
                    │  UserService: [modifying]   │
                    │  PayService:  [claimed]     │
                    │  CacheService: [available]  │
                    │                             │
                    └─────────────────────────────┘

Two Coordination Mechanisms

1. Pheromone System (Stigmergic)

Agents leave markers on code nodes that decay over time—like ants leaving scent trails:

Pheromone	Half-Life	Meaning
exploring	2 min	"I'm looking at this"
claiming	1 hour	"This is my territory"
modifying	10 min	"I'm actively changing this"
completed	24 hours	"I finished work here"
warning	Never	"Don't touch this"
blocked	5 min	"I'm stuck"

Self-healing: If an agent crashes, its pheromones decay and the work becomes available again.

2. Task Queue (Blackboard)

Explicit task management with dependencies:

┌─────────────────────────────────────────────────────────────┐
│                        TASK QUEUE                           │
├─────────────────────────────────────────────────────────────┤
│ [available] Add JSDoc to UserService         priority: high │
│ [claimed]   Add JSDoc to AuthService         agent: worker1 │
│ [blocked]   Update API docs ─────────────────► depends on ──┤
│ [in_progress] Add JSDoc to PaymentService    agent: worker2 │
│ [completed] Add JSDoc to CacheService        ✓              │
└─────────────────────────────────────────────────────────────┘

Swarm Tools

Tool	Purpose
swarm_post_task	Add a task to the queue
swarm_get_tasks	Query tasks with filters
swarm_claim_task	Claim/start/release a task
swarm_complete_task	Complete/fail/request review
swarm_pheromone	Leave a marker on a code node
swarm_sense	Query what other agents are doing
swarm_cleanup	Remove pheromones after completion

Example: Parallel Refactoring

// Orchestrator decomposes the task and creates individual work items
swarm_post_task({
  projectId: "backend",
  swarmId: "swarm_rename_user",
  title: "Update UserService.findUserById",
  description: "Rename getUserById to findUserById in UserService",
  type: "refactor",
  createdBy: "orchestrator"
})

// Workers claim and execute tasks
swarm_claim_task({ projectId: "backend", swarmId: "swarm_rename_user", agentId: "worker_1" })
// ... do work ...
swarm_complete_task({ taskId: "task_1", agentId: "worker_1", action: "complete", summary: "Renamed method" })

Install the Swarm Skill

For optimal swarm execution, install the included Claude Code skill that teaches agents the coordination protocol:

# Copy to your global skills directory
mkdir -p ~/.claude/skills
cp -r skills/swarm ~/.claude/skills/

Or for a specific project:

cp -r skills/swarm .claude/skills/

The skill provides:

• Worker agent protocol with step-by-step workflow
• Multi-phase orchestration patterns (discovery, contracts, implementation, validation)
• Common failure modes and how to prevent them
• Complete tool reference

Once installed, just say "swarm" or "parallel agents" and Claude will use the skill automatically.

See skills/swarm/SKILL.md for the full documentation.

---

All Tools

Tool	Description
Discovery
list_projects	List parsed projects in the database
search_codebase	Semantic search using vector embeddings
traverse_from_node	Explore relationships from a node
natural_language_to_cypher	Convert questions to Cypher queries
Analysis
impact_analysis	Assess refactoring risk (LOW/MEDIUM/HIGH/CRITICAL)
detect_dead_code	Find unreferenced exports and methods
detect_duplicate_code	Find structural and semantic duplicates
Parsing
parse_typescript_project	Build the graph from source
check_parse_status	Monitor async parsing jobs
start_watch_project	Auto-update graph on file changes
stop_watch_project	Stop file watching
list_watchers	List active file watchers
Swarm
swarm_post_task	Add task to the queue
swarm_get_tasks	Query tasks
swarm_claim_task	Claim/start/release tasks
swarm_complete_task	Complete/fail/review tasks
swarm_pheromone	Leave coordination markers
swarm_sense	Query what others are doing
swarm_cleanup	Clean up after swarm completion
Utility
test_neo4j_connection	Verify database connectivity

Tool Workflow Patterns

Pattern 1: Discovery → Focus → Deep Dive

list_projects → search_codebase → traverse_from_node → traverse (with skip for pagination)

Pattern 2: Pre-Refactoring Safety

search_codebase("function to change") → impact_analysis(nodeId) → review risk level

Pattern 3: Code Health Audit

detect_dead_code → detect_duplicate_code → prioritize cleanup

Pattern 4: Multi-Agent Work

swarm_post_task → swarm_claim_task → swarm_complete_task → swarm_get_tasks(includeStats) → swarm_cleanup

Multi-Project Support

All query tools require projectId for isolation. You can use:

• Project ID: proj_a1b2c3d4e5f6 (auto-generated)
• Project name: my-backend (from package.json)
• Project path: /path/to/project (resolved automatically)

// These all work:
search_codebase({ projectId: "my-backend", query: "auth" })
search_codebase({ projectId: "proj_a1b2c3d4e5f6", query: "auth" })
search_codebase({ projectId: "/path/to/my-backend", query: "auth" })

---

Framework Support

NestJS (Built-in)

Deep understanding of NestJS patterns:

• Controllers with route analysis (@Controller, @Get, @Post, etc.)
• Services with dependency injection mapping (@Injectable)
• Modules with import/export relationships (@Module)
• Guards, Pipes, Interceptors as middleware chains
• DTOs with validation decorators (@IsString, @IsEmail, etc.)
• Entities with TypeORM relationship mapping

NestJS-Specific Relationships:

• INJECTS - Dependency injection
• EXPOSES - Controller exposes HTTP endpoint
• MODULE_IMPORTS, MODULE_PROVIDES, MODULE_EXPORTS - Module system
• GUARDED_BY, TRANSFORMED_BY, INTERCEPTED_BY - Middleware

Custom Framework Schemas

The parser is config-driven. Define your own framework patterns:

// Example: Custom React schema
const REACT_SCHEMA = {
  name: 'react',
  decoratorPatterns: [
    { pattern: /^use[A-Z]/, semanticType: 'ReactHook' },
    { pattern: /^with[A-Z]/, semanticType: 'HOC' },
  ],
  nodeTypes: [
    { coreType: 'FunctionDeclaration', condition: (node) => node.name?.endsWith('Provider'), semanticType: 'ContextProvider' },
  ],
  relationships: [
    { type: 'PROVIDES_CONTEXT', from: 'ContextProvider', to: 'ReactHook' },
  ]
};

The dual-schema system means every node has:

• coreType: AST-level (ClassDeclaration, FunctionDeclaration)
• semanticType: Framework meaning (NestController, ReactHook)

This enables queries like "find all hooks that use context" while maintaining AST precision for refactoring.

---

Embedding Configuration

Local embeddings are the default — no API key needed. The Python sidecar starts automatically on first use and runs a local model for high-quality code embeddings.

The sidecar uses MPS (Apple Silicon GPU) when available, falling back to CPU. It auto-shuts down after 3 minutes of inactivity to free memory, and restarts lazily when needed (~15-20s).

Device override: Set EMBEDDING_DEVICE=cpu to force CPU if MPS causes issues.

Half precision: Set EMBEDDING_HALF_PRECISION=true to load the model in float16, roughly halving memory usage.

Available Models

Set via the EMBEDDING_MODEL environment variable:

Model	Dimensions	RAM (fp16)	Quality	Best For
codesage/codesage-base-v2 (default)	1024	~700 MB	Best	Default, code-specific encoder, fast
Qodo/Qodo-Embed-1-1.5B	1536	~4.5 GB	Great	Machines with 32+ GB RAM
BAAI/bge-base-en-v1.5	768	~250 MB	Good	General purpose, low RAM
sentence-transformers/all-MiniLM-L6-v2	384	~100 MB	OK	Minimal RAM, fast
nomic-ai/nomic-embed-text-v1.5	768	~300 MB	Good	Code + prose mixed
sentence-transformers/all-mpnet-base-v2	768	~250 MB	Good	Balanced quality/speed
BAAI/bge-small-en-v1.5	384	~65 MB	OK	Smallest footprint

Example: Use a lightweight model on a low-memory machine:

claude mcp add --scope user code-graph-context \
  -e EMBEDDING_MODEL=BAAI/bge-base-en-v1.5 \
  -- code-graph-context

Switching Models

Switching models requires re-parsing — vector index dimensions are locked per model. Drop existing indexes first:

DROP INDEX embedded_nodes_idx IF EXISTS;
DROP INDEX session_notes_idx IF EXISTS;

Then re-parse your project with the new model configured.

Using OpenAI Instead

If you prefer OpenAI embeddings (higher quality, requires API key):

claude mcp add --scope user code-graph-context \
  -e OPENAI_EMBEDDINGS_ENABLED=true \
  -e OPENAI_API_KEY=sk-your-key-here \
  -- code-graph-context

---

Troubleshooting

MCP Server Not Connecting

# Check the server is registered
claude mcp list

# Verify Neo4j is running
docker ps | grep neo4j

# Test manually
code-graph-context status

Embedding Errors

"Failed to generate embedding" — The local sidecar may not have started. Check:

# Verify Python deps are installed
code-graph-context status

# Re-run init to fix sidecar setup
code-graph-context init

Out of memory (large model on 16GB machine) — Switch to a lighter model:

claude mcp add --scope user code-graph-context \
  -e EMBEDDING_MODEL=BAAI/bge-base-en-v1.5 \
  -- code-graph-context

Using OpenAI and getting auth errors — Ensure your key is configured:

claude mcp remove code-graph-context
claude mcp add --scope user code-graph-context \
  -e OPENAI_EMBEDDINGS_ENABLED=true \
  -e OPENAI_API_KEY=sk-your-key-here \
  -- code-graph-context

Neo4j Memory Issues

For large codebases, increase memory limits:

# Stop and recreate with more memory
code-graph-context stop
code-graph-context init --memory 4G

Parsing Timeouts

Use async mode for large projects:

parse_typescript_project({
  projectPath: "/path/to/project",
  tsconfigPath: "/path/to/project/tsconfig.json",
  async: true  // Returns immediately, poll with check_parse_status
})

---

CLI Commands

code-graph-context init [options]   # Set up Neo4j + Python sidecar + embedding model
code-graph-context status           # Check Docker/Neo4j/sidecar status
code-graph-context stop             # Stop Neo4j container

Init options:

• -p, --port <port> - Bolt port (default: 7687)
• --http-port <port> - Browser port (default: 7474)
• --password <password> - Neo4j password (default: PASSWORD)
• -m, --memory <size> - Heap memory (default: 2G)
• -f, --force - Recreate container

---

Contributing

git clone https://github.com/andrew-hernandez-paragon/code-graph-context.git
cd code-graph-context
npm install
npm run build
npm run dev  # Watch mode

Conventional Commits: feat|fix|docs|refactor(scope): description

---

License

MIT - see LICENSE

---

Links

• Issues
• MCP Documentation
• Neo4j