ts-graph-mcp

A TypeScript code graph tool that extracts code structure into a queryable database, with MCP integration for AI coding agents.

What It Does

ts-graph parses TypeScript source code using AST analysis and builds a graph database of your codebase structure. The graph captures code symbols (functions, classes, interfaces, types, variables) and their relationships (calls, imports, type usage, inheritance).

Semantic search included. On first run, ts-graph downloads an embedding model (~300MB) and generates embeddings for all symbols. AI agents can search by concept ("user validation", "database queries") not just exact symbol names.

AI agents query the graph through the searchGraph MCP tool to:

• Find code by concept (semantic search)
• Traverse call graphs (who calls this? what does this call?)
• Find paths between symbols

Quick Start

1. Configuration

Create ts-graph-mcp.config.json in your project root:

{
  "packages": [
    { "name": "main", "tsconfig": "./tsconfig.json" }
  ],
  "server": {
    "port": 4000
  }
}

For monorepos, list multiple packages:

{
  "packages": [
    { "name": "shared", "tsconfig": "./shared/tsconfig.json" },
    { "name": "frontend", "tsconfig": "./frontend/tsconfig.json" },
    { "name": "backend", "tsconfig": "./backend/tsconfig.json" }
  ],
  "server": {
    "port": 4000
  }
}

2. Start the HTTP Server

npx ts-graph-mcp

The server indexes your project on first run and watches for changes.

3. Configure Claude Code

claude mcp add ts-graph-mcp -- npx -y ts-graph-mcp --mcp

Or manually in .mcp.json:

{
  "mcpServers": {
    "ts-graph-mcp": {
      "command": "npx",
      "args": ["-y", "ts-graph-mcp", "--mcp"]
    }
  }
}

Note: Start the HTTP server first. The MCP wrapper connects to it.

MCP Tool: searchGraph

Unified search combining semantic search with graph traversal.

Query Patterns

// Find code by concept (semantic search)
{ topic: "user authentication" }

// What does handleRequest call? (forward traversal)
{ from: { symbol: "handleRequest" } }

// Who calls saveUser? (backward traversal)
{ to: { symbol: "saveUser" } }

// How does A reach B? (path finding)
{ from: { symbol: "handleRequest" }, to: { symbol: "saveUser" } }

Parameters

Parameter	Required	Description
topic	No*	Standalone semantic search (not combinable with from/to)
from	No*	Start point: { symbol } or { query } with optional file_path
to	No*	End point: { symbol } or { query } with optional file_path
max_nodes	No	Output limit (default: 50)

*At least one of topic, from, or to is required. topic is standalone only — it cannot be combined with from/to.

Example Output

## Symbols matching "validation" (semantic search)

validateInput (Function) - src/validation.ts [score: 0.847]
checkUserData (Function) - src/user.ts [score: 0.721]

## Graph

handleRequest --CALLS--> validate --CALLS--> saveUser

## Nodes

validate:
  type: Function
  file: src/service.ts
  offset: 10, limit: 5
  snippet:
    10: export function validate(data: Input) {
  > 11:   return saveUser(data);
    12: }

CLI Options

ts-graph-mcp              # Start HTTP server
ts-graph-mcp --mcp        # Start MCP stdio server
ts-graph-mcp --reindex    # Force clean reindex

Configuration Reference

Required

Field	Description
packages	Array of { name, tsconfig }
server.port	HTTP server port (no default)

Optional

Field	Description	Default
embedding.preset	Embedding model	"nomic-embed-text-v1.5"
storage.type	Database type	"sqlite"
storage.path	Database file path	.ts-graph-mcp/graph.db
watch.debounce	Enable debouncing	true
watch.debounceInterval	Debounce delay (ms)	300
watch.polling	Use polling (for Docker/WSL2)	false
watch.pollingInterval	Polling interval (ms)	1000
watch.excludeDirectories	Directories to skip	[]
watch.silent	Suppress reindex logs	false

Embedding Models

Preset	Size	Dimensions	Notes
nomic-embed-text-v1.5	~300MB	768	Default, fast and effective
qwen3-0.6b	~650MB	1024	Higher quality, slower
qwen3-4b	~4GB	2560	Highest quality, needs more RAM
jina-embeddings-v2-base-code	~300MB	768	Optimized for code

Add .ts-graph-mcp/ to your .gitignore.

Yarn PnP Support

ts-graph works with Yarn 4 PnP monorepos. When .pnp.cjs is detected, module resolution uses Yarn's PnP API.

Requirements:

• Use base package imports (@libs/utils, not @libs/utils/date)
• Declare dependencies with workspace:* protocol

Supported Types

Nodes: Function, Class, Method, Interface, TypeAlias, Variable, SyntheticType, Feature, Spec, TestSuite, Test

Edges: CALLS, IMPLEMENTS, EXTENDS, TAKES, RETURNS, HAS_TYPE, HAS_PROPERTY, DERIVES_FROM, ALIAS_FOR, REFERENCES, INCLUDES, CONTAINS, SPECIFIES, VERIFIED_BY

Development

npm run check        # Run tests, build, and lint
npm test             # Run tests
npm run build        # Compile TypeScript

Project Structure

ts-graph-mcp/
├── http/          # HTTP server, database, ingestion, queries
├── mcp/           # MCP stdio wrapper
├── shared/        # Shared types
├── ui/            # Web UI (React + Vite)
└── main.ts        # Entry point

Windows Users

This package uses better-sqlite3, which requires compilation tools:

1. Install Visual Studio Build Tools with "Desktop development with C++"
2. Install Python 3.x
3. Use Node.js LTS

Contributing

See ARCHITECTURE.md for technical internals and CLAUDE.md for code style guidelines.

License

MIT