onboardai 0.1.3

An AI-powered onboarding intelligence layer on top of Graphify for MCP

OnboardAI: AI-Powered Onboarding Intelligence Layer

OnboardAI is a developer onboarding intelligence system that integrates with Claude Code and other MCP-compatible clients. It is designed to dramatically reduce the time it takes a new developer, intern, or employee to understand a codebase by converting raw repository data and Graphify knowledge graphs into interactive, onboarding-oriented insights.

---

1. Vision & Architecture

Rather than generating raw repository diagrams or acting as another codebase crawler, OnboardAI acts as a structured guidance layer on top of Graphify. It parses Graphify’s extraction data, maps relationships with NetworkX, groups code modules by directory structure, and answers questions relative to what tasks a developer has been assigned.

Core Architecture Flow

flowchart TD
    subgraph Client Layer
        ClaudeCode[Claude Code / MCP Client]
    end

    subgraph OnboardAI MCP Server
        Server[FastMCP Server]
        Scanner[Project Scanner & Tech Detector]
        Adapter[Graphify Adapter]
        Roadmap[Roadmap Generator]
        QA[Q&A & Context Engine]
        Viz[Mermaid Visualizer]
    end

    subgraph Intelligence Engine
        Graphify[Graphify extraction]
        NX[NetworkX Directed Graph]
    end

    ClaudeCode <-->|JSON-RPC via Stdio/SSE| Server
    Server <--> Scanner
    Server <--> Adapter
    Adapter <-->|Reads graphify-out/| Graphify
    Adapter <-->|Builds| NX
    Roadmap <--> Adapter
    QA <--> Adapter
    Viz <--> Adapter

---

2. Directory Layout & Components

Here is the file structure of the OnboardAI project:

OnboardAI/
├── .graphifyignore        # Ignore rules to bypass LLM extraction for non-code files
├── requirements.txt       # Project dependencies (FastAPI, uvicorn, mcp, networkx)
├── main.py                # Launcher: starts MCP server in stdio or sse mode
├── test_integration.py    # Offline self-check verification test script
├── test_client.py         # Standard python MCP client session simulation script
├── scanner/
│   ├── __init__.py
│   ├── project_scanner.py     # Counts files and crawls folders
│   └── architecture_detector.py # Detects tech stacks (FastAPI, React, etc.)
├── knowledge/
│   ├── __init__.py
│   ├── graphify_adapter.py    # Integrates Graphify output and builds NetworkX graphs
│   ├── onboarding_plan.py     # Generates day-by-day study paths
│   ├── repository_inventory.py# Structured module/class/function catalogs
│   └── qa_assistant.py        # Performs graph-based Q&A and task instructions
└── graph/
    ├── __init__.py
    └── graph_generator.py     # Generates Mermaid mindmaps, dependencies, and call-flows

---

3. How Each Feature Works & How It's Built

Feature 1: Repository Scan & Graph Integration

• How it's built: Implemented in scanner/project_scanner.py and scanner/architecture_detector.py.
• How it works: Walks the codebase directories (excluding noise like .git, node_modules, and virtual environments). It looks for configuration signatures (e.g. package.json, requirements.txt, manage.py, Dockerfile) and extracts details.
• Graphify Integration: Located in knowledge/graphify_adapter.py. If graphify-out/graph.json is missing, the adapter automatically invokes python -m graphify extract <repo_path> via a subprocess to bootstrap the graph. It then builds a networkx.DiGraph to represent the entities and relationships.

Feature 2: Dependency-Aware Learning Roadmap

• How it's built: Implemented in knowledge/onboarding_plan.py.
• How it works: Groups files using Graphify’s community detection. It builds a directed graph of community dependencies (if Community A calls/imports elements in Community B, then B is foundational to A). It uses topological sorting (or sorting by in/out-degree ratios) to order the learning path:
  • Day 1: Environment setup, tech summary, configs.
  • Day 2: Foundational modules (utilities, schemas, database clients).
  • Day 3: Mid-level business logic.
  • Day 4: API controllers, routers, and entry points.
  • Day 5: Automated tests and first contribution.

Feature 3: Mindmap & Dependency Visualization

• How it's built: Implemented in graph/graph_generator.py.
• How it works: Traversing the NetworkX graph, it extracts relationships and formats them directly into Mermaid syntax:
  • generate_mindmap(): Creates a Mermaid mindmap showing high-level folders and key classes.
  • generate_call_flow(): Creates a flow diagram (flowchart TD) of method-to-method calls.
  • generate_dependency_graph(): Creates an import diagram (flowchart LR) of module dependencies.

Feature 4: Interactive Q&A & Task-Context Overlays

• How it's built: Implemented in knowledge/qa_assistant.py.
• How it works: Uses regex and keywords to search the graph. If a query matches a module (e.g. ProjectScanner), it generates a structured analysis of the component: Purpose, Responsibilities, Methods, Dependencies, and Callers.
• Task-Context Overlays: If the client provides a work_context string (e.g., "I need to implement a visualizer"), the Q&A assistant searches for components associated with those keywords in the graph and appends tailored, step-by-step developer instructions.

---

4. MCP Server & Protocols

The MCP Server is implemented in mcp_server/server.py using the high-level FastMCP SDK. This enables automatic tool schemas, JSON-RPC communication, and transport selection.

It exposes 10 tools to any connected client:

1. analyze_repository(repo_path): High-level tech stack and file count summary.
2. generate_onboarding_plan(repo_path, role, team, target_module): Customized daily learning roadmap.
3. explain_project(repo_path): Basic codebase summary.
4. explain_module(repo_path, module_name): Full signature and relationship breakout for a module.
5. list_modules(repo_path): Direct directory list of files and components.
6. search_repository(repo_path, query): Search symbols/descriptions in the codebase.
7. get_architecture(repo_path, visualization_type): Generates Mermaid charts.
8. get_learning_path(repo_path): Retrieves recommended files to read first.
9. get_dependencies(repo_path): Returns internal couplings and external dependencies.
10. ask_repository_question(repo_path, question, work_context): Q&A helper with custom action items.

---

5. Setup & Integration Guide

Local Verification

Verify the server runs successfully on your machine by running the test client session simulation:

$env:PYTHONIOENCODING="utf-8"
python test_client.py

This script acts as a local client session, connects to the server, retrieves the tools list, and invokes them locally.

Integrating with Claude Code

To register the server with the Claude Code CLI, run:

claude mcp add onboard-ai python "C:\Users\RohitAnish\.gemini\antigravity\scratch\OnboardAI\main.py"

Once registered, launch Claude Code:

claude

Integrating with Claude Desktop

To integrate with the Claude Desktop App, open your desktop configuration file:

• Path: %APPDATA%\Claude\claude_desktop_config.json
• Configuration:

  {
    "mcpServers": {
      "onboard-ai": {
        "command": "python",
        "args": [
          "C:/Users/RohitAnish/.gemini/antigravity/scratch/OnboardAI/main.py"
        ]
      }
    }
  }
  

---

6. How to Ask Questions

Once integrated into Claude, the model will automatically pick the right tool based on your natural language prompt. Here are some examples:

Question you ask Claude	Tool Claude calls under the hood
"How should I learn this project?"	generate_onboarding_plan
"Explain the module ProjectScanner"	explain_module
"Which components are critical to look at?"	get_learning_path
"Show me the dependency mindmap"	get_architecture(visualization_type='mindmap')
"How does the project work? I am going to work on the visualizer."	ask_repository_question(work_context='visualizer')