⚡ cgraph

Your codebase, as a queryable graph.

Give AI agents instant structural awareness — callers, callees, impact analysis, trace paths — without scanning thousands of files.

  ┌─────────┐     MCP      ┌──────────┐    SQLite    ┌──────────┐
  │ Copilot │◄────────────►│  cgraph  │◄────────────►│ .cgraph/ │
  │  Agent  │  JSON-RPC    │  Server  │   sql.js     │ graph.db │
  └─────────┘              └──────────┘              └──────────┘
       │                        ▲                         ▲
       │ "who calls             │ 22 tools                │ files, nodes,
       │  handleRequest?"       │ instant response        │ edges, roles
       ▼                        │                         │
  ┌─────────┐              ┌──────────┐              ┌──────────┐
  │  1 call │  instead of  │  Parser  │───index────►│ Your Code│
  │  2.2s   │              │ + Walker │              │  .ts .py │
  └─────────┘              └──────────┘              │  .c  .sh │
                                                     └──────────┘

Getting Started · MCP Tools · CLI Reference · Benchmarks · Architecture

🎯 The Problem

AI agents waste 70% of tool calls on grep → read_file → grep → read_file chains just to understand code structure. Every question triggers a cascade:

❌ Without cgraph                          ✅ With cgraph
─────────────────────────────              ──────────────────────
1. grep "handleRequest"                    1. cgraph_node handleRequest
2. read_file server.ts                        → definition, callers,
3. grep "import.*handleRequest"               callees, file:line
4. read_file routes.ts                        all in ONE response
5. grep "routes" to find callers
6. read_file app.ts
7. finally has the answer                  Done. 2.2 seconds.
   14 seconds later...

cgraph pre-computes the graph once, then answers structural queries instantly.

✨ Feature Highlights

🔌 Install & Forget

One-command installer. Auto-indexes on first Copilot query. No config per project.

🛠️ 22 MCP Tools

search · context · trace · explore · node · callers · callees · impact · files · status · affected · export · changed · deadcode · cycles · stats · suggest · auto-context · intent-search · validate-plan · lint · dna

⚡ Dramatically Faster Workflows

Collapses multi-step grep→read chains into single precomputed graph queries. Avg MCP tool latency: 17ms.

🌍 Multi-Language

TypeScript · JavaScript · Python · C · C++ · Shell · PowerShell — all from one index.

🧠 Smart Analysis

Dead code detection · cycle finding · refactoring suggestions · role classification · project statistics.

🔄 Incremental & Live

3-tier change detection. File watcher with auto re-index. Parallel parsing with worker threads.

🚀 Getting Started

Option A: One-Command Install

No Node.js or git required — the installer handles everything.

Windows

powershell -ExecutionPolicy Bypass -File install.ps1

macOS / Linux

bash install.sh

The installer will:

📦 Download portable Node.js (if needed)
📥 Fetch cgraph source
🔨 Build it
🔗 Add cgraph to your PATH
⚙️ Configure VS Code MCP globally (works in all workspaces)

That's it. Open any project → ask Copilot → cgraph auto-indexes and responds.

Option B: From Source

git clone https://github.com/samarth-w/agent_graph.git
cd agent_graph
npm install && npm run build
npm link                       # optional: makes `cgraph` available globally

📋 VS Code MCP Configuration (if you installed from source)

Add to your VS Code settings.json (Ctrl+Shift+P → "Open User Settings (JSON)"):

{
  "mcp": {
    "servers": {
      "cgraph": {
        "command": "node",
        "args": ["<path-to-cgraph>/bin/cgraph.js", "serve", "--mcp"],
        "cwd": "${workspaceFolder}"
      }
    }
  }
}

Windows tip: If VS Code can't find node, use "C:\\Program Files\\nodejs\\node.exe".

👥 Teammate Setup

With Node.js: git clone → npm install → npm run build → npm link → add MCP config.

Without Node.js: Share the repo folder → run install.ps1 (Win) or install.sh (Mac/Linux) → done.

🔧 MCP Tools

When running as an MCP server, 22 tools are available to AI agents:

graph LR
    subgraph "🔍 Discovery"
        A[cgraph_search] --> B[cgraph_explore]
        A --> C[cgraph_node]
    end
    subgraph "🧭 Navigation"
        C --> D[cgraph_callers]
        C --> E[cgraph_callees]
        D --> F[cgraph_trace]
    end
    subgraph "💡 Intelligence"
        F --> G[cgraph_context]
        G --> H[cgraph_impact]
        H --> I[cgraph_affected]
    end
    subgraph "📊 Analysis"
        J[cgraph_deadcode]
        K[cgraph_cycles]
        L[cgraph_stats]
        M[cgraph_suggest]
    end
    subgraph "🧠 Agentic Intelligence"
        R[cgraph_auto_context]
        S[cgraph_intent_search]
        T[cgraph_validate_plan]
        U[cgraph_lint]
        V[cgraph_dna]
    end
    subgraph "📦 Utilities"
        N[cgraph_status]
        O[cgraph_files]
        P[cgraph_export]
        Q[cgraph_changed]
    end

    style A fill:#3B82F6,color:#fff
    style G fill:#8B5CF6,color:#fff
    style H fill:#EF4444,color:#fff
    style M fill:#F59E0B,color:#fff
    style R fill:#10B981,color:#fff
    style T fill:#EC4899,color:#fff

Tool	What it does	When to use
`cgraph_context`	Builds ranked code context for a task	Start here — best for architecture & feature questions
`cgraph_search`	Find symbols by name with fuzzy matching	Looking for a specific function or class
`cgraph_node`	Symbol detail + full call trail	Deep-dive on one symbol
`cgraph_explore`	Source code for multiple related symbols	Need actual code, not just structure
`cgraph_callers`	Who calls this? (reverse graph)	Understanding usage patterns
`cgraph_callees`	What does this call? (forward graph)	Understanding dependencies
`cgraph_trace`	Call path between two symbols	"How does X reach Y?"
`cgraph_impact`	Blast radius of a change	Pre-change risk assessment
`cgraph_affected`	Test files impacted by changes	CI optimization, test selection
`cgraph_changed`	Symbols changed in git diff	Code review, change mapping
`cgraph_deadcode`	Unreachable symbols	Cleanup candidates
`cgraph_cycles`	Circular dependency detection	Architecture health
`cgraph_stats`	Project metrics & hotspots	Codebase overview
`cgraph_suggest`	Refactoring suggestions	Extract, inline, move, split recommendations
`cgraph_export`	Mermaid / DOT / HTML diagrams	Visualization & docs
`cgraph_files`	List all indexed files	Inventory check
`cgraph_auto_context`	File-level warm start: symbols, callers, callees, tests	Open a file — instant awareness before coding
`cgraph_intent_search`	Natural language symbol search (BM25)	"find auth middleware" — searches by meaning, not just name
`cgraph_validate_plan`	Pre-flight change risk assessment	Before refactoring — blast radius, affected tests, risk score
`cgraph_lint`	Architecture rule enforcement	CI gate — deny imports, max fan-out, cycle checks
`cgraph_dna`	Codebase fingerprint & health scores	Onboarding — languages, architecture style, health overview
`cgraph_status`	Index health & stats	Debugging, verification

💻 CLI Reference

cgraph <command> [options]

Core Commands

Command	Description
`index [dir]`	Build / update the code graph (incremental)
`sync [dir]`	Re-index changed files only
`search <query>`	Search symbols — supports `kind:`, `lang:`, `path:`, `role:`, `exported:` filters
`callers <symbol>`	Reverse call graph — who calls this?
`callees <symbol>`	Forward call graph — what does this call?
`impact <symbol>`	What breaks if this changes?
`trace <from> <to>`	Find the call path between two symbols
`context <task>`	Build ranked code context for a task description
`explore <query>`	Get source code for related symbols
`node <symbol>`	Symbol detail with call trail
`query <symbol>`	Look up a symbol with callers/callees
`where <symbol>`	Find where a symbol is defined

Analysis Commands

Command	Description
`deadcode`	Find unreachable symbols (dead code)
`cycles`	Detect circular dependencies
`stats`	Project metrics — hotspots, coupling, complexity
`suggest`	AI-powered refactoring suggestions

Agentic Intelligence Commands

Command	Description
`auto-context <file>`	File-level warm start — symbols, callers, callees, related tests
`intent <query>`	Natural language symbol search (BM25 scoring)
`validate`	Pre-flight change risk assessment from stdin
`lint`	Architecture rule enforcement via `.cgraph.json` rules
`dna`	Codebase fingerprint — languages, health scores, architecture style

Infrastructure Commands

Command	Description
`status`	Index health (files, nodes, edges, languages, roles)
`files`	List all indexed files
`affected <files>`	Find test files impacted by changes
`export`	Generate Mermaid, DOT, or interactive HTML diagrams
`changed`	Map git diff to changed symbols
`watch [dir]`	Watch for file changes and auto re-index
`serve --mcp`	Start MCP server (JSON-RPC 2.0 over stdio)

Global Options

--depth <n>       Max traversal depth (default: 3)
--max-nodes <n>   Max nodes to return (default: 50)
--kind <kind>     Filter by symbol kind (function, class, method, etc.)
--file <path>     Filter by file path
--json            Raw JSON output
--pretty          Formatted output

Project Configuration

Drop a .cgraph.json in your project root to customize behavior:

{
  "maxDepth": 5,
  "maxNodes": 100,
  "ignorePaths": ["vendor", "generated"],
  "extensions": [".ts", ".tsx", ".py"]
}

📊 Benchmarks

Benchmarked on a real-world TypeScript finance app — 16 files, 168 symbols, 1,006 edges. MCP tool latency benchmarked on cgraph's own codebase — 81 files, 687 nodes, 1,282 edges.

Agent Question	Without cgraph	With cgraph	Speedup
Where is `formatMoney` defined and who calls it?	7 calls · 14.0s	1 call · 2.2s	🟢 6.5x
Impact of changing `ApiController`?	6 calls · 12.0s	1 call · 2.2s	🟢 5.6x
How does `createApp` reach `formatMoney`?	4 calls · 8.0s	1 call · 2.2s	🟢 3.6x
Changed `store.ts` — what tests to run?	5 calls · 10.0s	1 call · 2.2s	🟢 4.6x
Explain the architecture	11 calls · 22.0s	2 calls · 4.4s	🟢 5.0x
Total	33 calls · 66s	6 calls · 13s	⚡ 5.1x faster

MCP Tool Latency (self-hosted benchmark)

Metric	Value
Avg latency (22 tools)	17ms
Min latency	2ms (search, callees)
Max latency	103ms (explore — includes source read)
Cold index + query	566ms (auto-index on first call)
Warm re-sync	288ms (no-change check)
Burst (10× search)	6ms/call

Performance powered by bulk adjacency maps — getFileMap, getNodeMap, getAdjacencyMaps load the graph in 3 queries, then all lookups are O(1) map gets. Zero N+1 query patterns.

🏃 Run benchmarks yourself

# Agent workflow comparison (with vs without cgraph)
node scripts/benchmark-agent.mjs <your-project-dir>

# MCP server latency (22 tools, burst, cold start)
node scripts/benchmark.mjs

# Raw efficiency comparison (grep+read vs cgraph)
node scripts/benchmark-compare.mjs

🏗️ Architecture

How It Works

flowchart TB
    subgraph Index["📥 Indexing Pipeline"]
        direction TB
        W[File Walker] -->|"*.ts *.py *.c *.sh"| P[Parser]
        P -->|"symbols + calls"| S[Synthesizer]
        S -->|"dynamic dispatch edges"| R[Edge Resolver]
        R -->|"import-aware resolution"| C[Role Classifier]
    end

    subgraph Store["💾 Storage"]
        DB[(SQLite via sql.js)]
        DB --- F[files]
        DB --- N[nodes]
        DB --- E[edges]
    end

    subgraph Serve["🔌 MCP Server"]
        MCP[JSON-RPC 2.0] --> Cache[LRU Cache]
        Cache --> Q[Query Engine]
        Q --> BFS[BFS Traversal]
        Q --> CTX[Context Builder]
        Q --> ANA[Analysis Engine]
    end

    Index --> DB
    DB --> Serve
    Agent[🤖 AI Agent] <-->|stdio| MCP

    style Agent fill:#8B5CF6,color:#fff
    style DB fill:#3B82F6,color:#fff
    style MCP fill:#22C55E,color:#fff

Design Principles

Principle	Implementation
Zero cloud dependencies	Pure JS/WASM — `sql.js` instead of `better-sqlite3`, no native bindings
Incremental by default	3-tier: mtime+size → content hash → parse. Rebuilds only what changed
Import-aware resolution	Resolves calls through imports: import match > same-file > global name
Smart role classification	Symbols tagged as `entry` · `core` · `hub` · `bridge` · `utility` · `leaf` · `test` · `dead`
Bounded traversal	BFS with `maxDepth` + `maxNodes` caps + cycle detection
Token-conscious	Context payloads include estimated token counts so agents can budget
Bulk query optimization	Adjacency maps loaded in 3 SQL queries — all per-node lookups are O(1) map gets

Supported Languages

JavaScript `.js` `.jsx` `.mjs` `.cjs` _{via @babel/parser}	TypeScript `.ts` `.tsx` _{via @babel/parser}	Python `.py` `.pyi` _regex-based
C / C++ `.c` `.h` `.cpp` `.cc` `.hpp` _{regex-based · #include resolution}	Shell `.sh` `.bash` `.zsh` _{regex-based · source/. imports}	PowerShell `.ps1` `.psm1` `.psd1` _{regex-based · dot-sourcing}

Framework Detection

Route and endpoint extraction for: Express · React Router · Next.js · Flask · FastAPI · Django

Database Schema

Per-project SQLite at .cgraph/graph.db:

┌──────────┐     ┌──────────────────────────────────────────────┐     ┌───────────┐
│  files   │     │                   nodes                     │     │   edges   │
├──────────┤     ├──────────────────────────────────────────────┤     ├───────────┤
│ id       │◄───┐│ id · file_id · name · qualified_name        │┌───►│ source_id │
│ path     │    ││ kind · start_line · end_line · signature    ││    │ target_id │
│ hash     │    └┤ doc · exported · role                       ├┘    │ kind      │
│ language │     └──────────────────────────────────────────────┘     └───────────┘
│ mtime    │                        ▲
│ size     │     ┌──────────┐       │         ┌──────────┐
└──────────┘     │ raw_refs │───────┘         │ metadata │
                 │ caller   │                 │ key      │
                 │ callee   │                 │ value    │
                 │ kind     │                 └──────────┘
                 └──────────┘

📁 Project Structure

cgraph/
├── bin/cgraph.js                  # CLI entry point
├── src/
│   ├── cli.ts                     # 22 CLI commands (commander)
│   ├── config.ts                  # Configuration + .cgraph.json loader
│   ├── context.ts                 # Context builder (search → expand → snippets)
│   ├── graph.ts                   # BFS traversal, impact, trace, dead code, cycles, suggest
│   ├── indexer.ts                 # File walker + parallel parser + incremental edge resolver
│   ├── mcp.ts                     # MCP server (22 tools, JSON-RPC 2.0, progress notifications)
│   ├── storage.ts                 # GraphDB (sql.js WASM SQLite)
│   ├── parser.ts                  # Multi-language parser (babel + regex)
│   ├── synthesizer.ts             # Dynamic dispatch edge synthesis
│   ├── frameworks.ts              # Framework route extraction
│   ├── lint.ts                    # Architecture rule enforcement engine
│   ├── search.ts                  # Fuzzy symbol search + BM25 intent search
│   ├── export.ts                  # Mermaid / DOT / HTML diagram generation
│   ├── cache.ts                   # LRU cache with disk persistence
│   ├── watcher.ts                 # File watcher with debounced re-index
│   ├── git.ts                     # Git diff → changed symbol mapping
│   ├── adaptive.ts                # Dynamic traversal limits
│   ├── query-parser.ts            # Search query field extraction
│   ├── gitignore.ts               # .gitignore parsing
│   └── types.ts                   # All type definitions
├── __tests__/                     # 216 tests (vitest)
├── scripts/                       # Benchmarks, installers, smoke tests
├── demo/                          # Finance tracker + C++/Shell demos
├── install.ps1 / install.sh       # Standalone installers
└── .cgraph.json                   # Project-level configuration

🧪 Development

npm run build              # compile TypeScript
npm run dev                # watch mode (rebuild on save)
npm test                   # run 216 unit tests
npm run test:watch         # watch tests

📜 All scripts

Script	Purpose
`install.ps1` / `install.sh`	Standalone installer (downloads Node if needed)
`scripts/local-install.ps1` / `.sh`	Build + npm link for dev testing
`scripts/smoke-test.ps1` / `.sh`	22 end-to-end CLI tests
`scripts/setup-mcp.ps1`	Auto-configure MCP for a project
`scripts/benchmark.mjs`	MCP server latency (22 tools, burst, cold start)
`scripts/benchmark-agent.mjs`	Agent workflow benchmark — with vs without cgraph
`scripts/benchmark-compare.mjs`	Raw efficiency comparison (grep+read vs cgraph)

📚 Usage as Library

import { GraphDB } from 'cgraph/storage';
import { indexProject } from 'cgraph/indexer';
import { findCallers, suggestRefactorings } from 'cgraph/graph';
import { buildContext } from 'cgraph/context';

// Index & query
const db = await GraphDB.open('.cgraph/graph.db');
await indexProject('.');

const callers = findCallers(db, 'handleRequest', { maxDepth: 3 });
const suggestions = suggestRefactorings(db, { file: 'src/app.ts' });

db.close();

📄 License

AGPL-3.0 — see LICENSE for details.

For proprietary/closed-source usage, contact the author for a commercial license.

_{Built with ❤️ for developers who want their AI agents to actually understand their code.}

Name		Name	Last commit message	Last commit date
.github		.github
.vscode		.vscode
__tests__		__tests__
bin		bin
demo		demo
python_test		python_test
scripts		scripts
src		src
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
LICENSE		LICENSE
README.md		README.md
engine_promax.ipynb		engine_promax.ipynb
idea.md		idea.md
install.ps1		install.ps1
install.sh		install.sh
package-lock.json		package-lock.json
package.json		package.json
tsconfig.json		tsconfig.json
vitest.config.ts		vitest.config.ts

Navigation Menu

Repository files navigation

⚡ cgraph

Your codebase, as a queryable graph.

🎯 The Problem

✨ Feature Highlights

🔌 Install & Forget

🛠️ 22 MCP Tools

⚡ Dramatically Faster Workflows

🌍 Multi-Language

🧠 Smart Analysis

🔄 Incremental & Live

🚀 Getting Started

Option A: One-Command Install

Option B: From Source

🔧 MCP Tools

💻 CLI Reference

Core Commands

Analysis Commands

Agentic Intelligence Commands

Infrastructure Commands

Global Options

Project Configuration

📊 Benchmarks

MCP Tool Latency (self-hosted benchmark)

🏗️ Architecture

How It Works

Design Principles

Supported Languages

Framework Detection

Database Schema

📁 Project Structure

🧪 Development

📚 Usage as Library

📄 License

About

Releases

Packages

Contributors

Languages

Latest commit

History

Folders and files

Repository files navigation

⚡ cgraph

Your codebase, as a queryable graph.

🎯 The Problem

✨ Feature Highlights

🔌 Install & Forget

🛠️ 22 MCP Tools

⚡ Dramatically Faster Workflows

🌍 Multi-Language

🧠 Smart Analysis

🔄 Incremental & Live

🚀 Getting Started

Option A: One-Command Install

Option B: From Source

🔧 MCP Tools

💻 CLI Reference

Core Commands

Analysis Commands

Agentic Intelligence Commands

Infrastructure Commands

Global Options

Project Configuration

📊 Benchmarks

MCP Tool Latency (self-hosted benchmark)

🏗️ Architecture

How It Works

Design Principles

Supported Languages

Framework Detection

Database Schema

📁 Project Structure

🧪 Development

📚 Usage as Library

📄 License

About

Resources

Stars

Watchers

Forks

Releases

Packages

Contributors

Languages