ClawGraph: OpenClaw & NemoClaw Interactive Graph

AI-powered knowledge graph builder and RAG system for the OpenClaw and NemoClaw ecosystems

ClawGraph automatically crawls open-source repositories, extracts entities and relationships into a knowledge graph, and serves grounded answers via RAG — all secured with a 5-layer prompt injection defense. The interface is a Telegram app bot that uses the OpenClaw skill to interact with the knowledge graph.

Conceptual Components

Conceptual Component	Implementation
Orchestration Pipeline	Directed Acyclic Graph-based crawl→extract→embed→graph→curate pipeline with retry and scheduling
Custom MCP Server	Python MCP server exposing 9 GitHub API tools via the Model Context Protocol
AI Pipeline	Gemini Flash Lite (extraction/classification) + Gemini 2.5 Flash (reasoning/curation)
Knowledge Graph & RAG	Neo4j / NetworkX graph + embedding search → graph-grounded answer generation with Chain of Thought reasoning
Prompt Injection Prevention	5-layer defense: sanitizer → classifier → canary tokens → output guardrails → audit

Architecture

GitHub Repos ──▶ [GitHub MCP Server] ──▶ [Orchestration Pipeline] ──▶ [Knowledge Graph]
                  (Python)                crawl→extract→embed→         (Neo4j / NetworkX)
                                          graph_update→curate                │
                                                                             ▼
Telegram Bot ◀── [OpenClaw Skill] ◀── [RAG Engine] ◀── [Graph + Vector Retrieval]
                                        (Gemini 2.5 Flash + CoT)
                                             │
                                    [5-Layer Security Defense]
                                    L1: Input Sanitizer
                                    L2: Injection Classifier (Flash Lite)
                                    L3: Canary Tokens
                                    L4: Output Guardrails
                                    L5: Audit Logger

Quick Start

Prerequisites

• Python 3.12+
• Gemini API key (free tier)
• GitHub PAT (no scope needed for public repos)
• Docker (optional, for containerized deployment)

Install

git clone https://github.com/YOUR_USERNAME/ClawGraph.git
cd ClawGraph
pip install -e ".[dev]"
cp .env.example .env
# Edit .env with your API keys

Run Tests

python -m pytest tests/ -v

Start the Server

python -m ClawGraph.main
# API available at http://localhost:8000

Run the Pipeline

# Via API
curl -X POST http://localhost:8000/api/pipeline/run

# Query the knowledge graph
curl -X POST http://localhost:8000/api/query \
  -H "Content-Type: application/json" \
  -d '{"question": "What is the Gateway in OpenClaw?"}'

Docker Deployment

docker compose build
docker compose up -d
# Health check: curl http://localhost:8000/api/health

📁 Project Structure

ClawGraph/
├── github_mcp_server/    # Custom Python MCP Server (9 GitHub tools)
├── pipeline/             # Orchestration Pipeline (5 stages + scheduler)
│   └── stages/           # crawl, extract, embed, graph_update, curate
├── graph/                # Knowledge Graph (Neo4j + NetworkX backends)
├── RAG/                  # RAG Engine (retriever, generator, embeddings)
├── security/             # 5-Layer Prompt Injection Defense
├── config.py             # 12-factor app configuration
├── models.py             # Pydantic data models
└── main.py               # FastAPI application
openclaw_skill/           # OpenClaw integration (SKILL.md + tool)
tests/                    # Tests including red-team security suite

Security Model

All user queries pass through a 5-layer defense pipeline:

1. L1 — Input Sanitizer: Strips injection delimiters, Unicode confusables, control characters
2. L2 — Injection Classifier: Gemini Flash Lite classifies input as benign/suspicious/malicious
3. L3 — Canary Tokens: Hidden UUID tokens in system prompts detect prompt leaks
4. L4 — Output Guardrails: Blocks system prompt fragments, credentials, and code execution patterns
5. L5 — Audit Logger: JSON-lines log of all security events for monitoring

GitHub MCP Server

Custom Python MCP server with 9 tools:

Tool	Description
get_repo_info	Repository metadata (stars, forks, language, topics)
list_repo_files	File tree with types and sizes
get_file_content	Raw file content (base64-decoded)
search_code	Code search across repositories
list_issues	Issues with labels and comments
list_pull_requests	PRs with merge status
list_forks	Forks sorted by stars/activity
get_commit_history	Recent commits with messages
get_contributors	Contributors with commit counts

Run standalone: python -m ClawGraph.github_mcp_server.server

📊 API Endpoints

Method	Endpoint	Description
POST	/api/query	RAG query with prompt injection defense
GET	/api/graph/stats	Knowledge graph statistics
POST	/api/pipeline/run	Trigger manual pipeline run
GET	/api/security/audit	Recent security events
GET	/api/health	Service health check

🧩 OpenClaw Integration

Install as an OpenClaw skill to use via Telegram:

/kg query What is the Gateway in OpenClaw?
/kg status
/kg crawl
/kg security-report

⚙️ Configuration

All configuration via environment variables (12-factor app):

Variable	Required	Default	Description
GEMINI_API_KEY	Yes	—	Google AI Studio API key
GITHUB_TOKEN	Yes	—	GitHub PAT
GRAPH_BACKEND	No	memory	neo4j or memory
NEO4J_URI	If neo4j	—	Neo4j Aura connection URI
PIPELINE_SCHEDULE	No	0 3 * * *	Cron schedule for auto-crawl
PIPELINE_TARGETS	No	openclaw/openclaw,NVIDIA/NemoClaw	Repos to crawl

📄 License

MIT