ARBuilder

AI-powered development assistant for the Arbitrum ecosystem. ARBuilder transforms natural language prompts into:

• Stylus smart contracts (Rust)
• Cross-chain SDK implementations (asset bridging and messaging)
• Full-stack dApps (contracts + backend + indexer + oracle + frontend + wallet integration)
• Orbit chain deployment assistance

Architecture

ARBuilder uses a Retrieval-Augmented Generation (RAG) pipeline with hybrid search (vector + BM25 + cross-encoder reranking) to provide context-aware code generation. Available as a hosted service at arbuilder.app or self-hosted via MCP server.

┌─────────────────────────────────────────────────────────────────────────┐
│                            ARBuilder                                    │
├─────────────────────────────────────────────────────────────────────────┤
│                                                                         │
│  DATA PIPELINE                                                          │
│  ┌──────────┐    ┌──────────┐    ┌───────────┐    ┌──────────────────┐ │
│  │ Scraper  │───▶│Processor │───▶│ Embedder  │───▶│    ChromaDB      │ │
│  │ crawl4ai │    │ 3-layer  │    │ BGE-M3    │    │ (local vectors)  │ │
│  │ + GitHub │    │ filters  │    │ 1024-dim  │    │                  │ │
│  └──────────┘    └──────────┘    └───────────┘    └────────┬─────────┘ │
│                                                             │           │
│  RETRIEVAL                                                  │           │
│  ┌──────────────────────────────────────────────────────────▼─────────┐ │
│  │                    Hybrid Search Engine                             │ │
│  │  ┌──────────┐    ┌──────────┐    ┌───────────┐                    │ │
│  │  │  Vector  │    │   BM25   │    │CrossEncoder│   RRF Fusion      │ │
│  │  │  Search  │───▶│ Keywords │───▶│ Reranker  │──▶ + MMR           │ │
│  │  └──────────┘    └──────────┘    └───────────┘                    │ │
│  └───────────────────────────────────────────────────────────────────┘ │
│                                         │                               │
│  GENERATION                             ▼                               │
│  ┌───────────────────────────────────────────────────────────────────┐  │
│  │                      MCP Server (14 tools)                        │  │
│  │                                                                   │  │
│  │  M1: Stylus        M2: SDK           M3: dApp Builder            │  │
│  │  ┌─────────────┐   ┌─────────────┐   ┌──────────────────────┐   │  │
│  │  │ generate_    │   │ generate_   │   │ generate_backend     │   │  │
│  │  │ stylus_code  │   │ bridge_code │   │ generate_frontend    │   │  │
│  │  │ ask_stylus   │   │ generate_   │   │ generate_indexer     │   │  │
│  │  │ get_context  │   │ messaging   │   │ generate_oracle      │   │  │
│  │  │ gen_tests    │   │ ask_bridging│   │ orchestrate_dapp     │   │  │
│  │  │ get_workflow │   │             │   │                      │   │  │
│  │  │ validate_code│   │             │   │                      │   │  │
│  │  └─────────────┘   └─────────────┘   └──────────────────────┘   │  │
│  └───────────────────────────────────────────────────────────────────┘  │
│                           │                                             │
│  IDE INTEGRATION          ▼                                             │
│  ┌───────────────────────────────────────────────────────────────────┐  │
│  │  Cursor / VS Code / Claude Desktop / Any MCP Client              │  │
│  │  <- via local stdio or remote mcp-remote proxy ->                │  │
│  └───────────────────────────────────────────────────────────────────┘  │
│                                                                         │
│  HOSTED SERVICE (Cloudflare Workers)                                    │
│  ┌──────────────┐  ┌──────────┐  ┌──────────┐  ┌──────────────────┐  │
│  │  Workers AI  │  │ Vectorize│  │    D1    │  │      KV         │  │
│  │  BGE-M3 +    │  │ 1024-dim │  │  Users   │  │   Source registry│  │
│  │  Reranker    │  │  index   │  │  API keys│  │   + Ingest state │  │
│  └──────────────┘  └──────────┘  └──────────┘  └──────────────────┘  │
│                                                                       │
│  INGESTION PIPELINE (Worker-native, cron every 6h)                    │
│  ┌──────────┐    ┌──────────┐    ┌───────────┐    ┌──────────────┐  │
│  │ scraper  │───▶│ chunker  │───▶│ Workers AI│───▶│  Vectorize   │  │
│  │ HTML/    │    │ doc+code │    │  BGE-M3   │    │   upsert     │  │
│  │ GitHub   │    │ splitter │    │ embedding │    │              │  │
│  └──────────┘    └──────────┘    └───────────┘    └──────────────┘  │
│                       │                ▲                             │
│                       │ >30 files      │ embed messages              │
│                       ▼                │                             │
│              ┌─────────────────────────┴──┐                         │
│              │    CF Queue (async path)    │                         │
│              │  embed │ continue │finalize │                         │
│              └────────────────────────────┘                         │
└─────────────────────────────────────────────────────────────────────────┘

TL;DR - Quick Start

Option 1: Hosted Service (Easiest)

# No local setup needed - just configure your IDE
# Add to ~/.cursor/mcp.json:
{
  "mcpServers": {
    "arbbuilder": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://arbuilder.app/mcp",
               "--header", "Authorization: Bearer YOUR_API_KEY"]
    }
  }
}

Get your API key at arbuilder.app

Option 2: Self-Hosted

# 1. Clone and setup
git clone https://github.com/Quantum3-Labs/ARBuilder.git
cd ARBuilder
conda env create -f environment.yml
conda activate arbbuilder

# 2. Configure API key
cp .env.example .env
# Edit .env and add your OPENROUTER_API_KEY and NVIDIA_API_KEY

# 3. Generate vector database (required)
python -m src.embeddings.vectordb

# 4. Test MCP server
python -m src.mcp.server
# Should show: "Capabilities: 14 tools, 11 resources, 5 prompts"

# 5. Configure Cursor IDE (~/.cursor/mcp.json) - see Setup section below

Tutorial Video

Watch the tutorial to see ARBuilder in action:

Tutorial Video

Project Structure

ArbBuilder/
├── sources.json          # Single source of truth for all data sources
├── scraper/              # Data collection module
│   ├── config.py         # Thin wrapper around sources.json (backward-compat helpers)
│   ├── scraper.py        # Web scraping with crawl4ai
│   ├── github_scraper.py # GitHub repository cloning
│   └── run.py            # Pipeline entry point
├── src/
│   ├── preprocessing/    # Text cleaning and chunking
│   │   ├── cleaner.py    # Text normalization
│   │   ├── chunker.py    # Document chunking with token limits
│   │   └── processor.py  # Main preprocessing pipeline
│   ├── embeddings/       # Embedding and vector storage
│   │   ├── embedder.py   # OpenRouter embedding client
│   │   ├── vectordb.py   # ChromaDB wrapper with hybrid search (BM25 + vector)
│   │   └── reranker.py   # CrossEncoder, MMR, LLM reranking
│   ├── templates/        # Code generation templates
│   │   ├── stylus_templates.py   # M1: Stylus contract templates
│   │   ├── backend_templates.py  # M3: NestJS/Express templates
│   │   ├── frontend_templates.py # M3: Next.js + wagmi templates
│   │   ├── indexer_templates.py  # M3: Subgraph templates
│   │   └── oracle_templates.py   # M3: Chainlink templates
│   ├── utils/            # Shared utilities
│   │   ├── version_manager.py   # SDK version management
│   │   ├── env_config.py        # Centralized env var configuration
│   │   ├── abi_extractor.py     # Stylus ABI extraction from Rust code
│   │   └── compiler_verifier.py # Docker-based cargo check verification
│   ├── mcp/              # MCP server for IDE integration
│   │   ├── server.py     # MCP server (tools, resources, prompts)
│   │   ├── tools/        # MCP tool implementations (14 tools)
│   │   │   ├── get_stylus_context.py   # M1
│   │   │   ├── generate_stylus_code.py # M1
│   │   │   ├── ask_stylus.py           # M1
│   │   │   ├── generate_tests.py       # M1
│   │   │   ├── get_workflow.py         # M1
│   │   │   ├── validate_stylus_code.py # M1
│   │   │   ├── generate_bridge_code.py # M2
│   │   │   ├── generate_messaging_code.py # M2
│   │   │   ├── ask_bridging.py         # M2
│   │   │   ├── generate_backend.py     # M3
│   │   │   ├── generate_frontend.py    # M3
│   │   │   ├── generate_indexer.py     # M3
│   │   │   ├── generate_oracle.py      # M3
│   │   │   └── orchestrate_dapp.py     # M3
│   │   ├── resources/    # Static knowledge (11 resources)
│   │   │   ├── stylus_cli.py      # M1
│   │   │   ├── workflows.py       # M1
│   │   │   ├── networks.py        # M1
│   │   │   ├── coding_rules.py    # M1
│   │   │   ├── sdk_rules.py       # M2
│   │   │   ├── backend_rules.py   # M3
│   │   │   ├── frontend_rules.py  # M3
│   │   │   ├── indexer_rules.py   # M3
│   │   │   └── oracle_rules.py    # M3
│   │   └── prompts/      # Workflow templates
│   └── rag/              # RAG pipeline (TBD)
├── tests/
│   ├── mcp_tools/        # MCP tool test cases and benchmarks
│   │   ├── test_get_stylus_context.py
│   │   ├── test_generate_stylus_code.py
│   │   ├── test_ask_stylus.py
│   │   ├── test_generate_tests.py
│   │   ├── test_m2_e2e.py    # M2 end-to-end tests
│   │   ├── test_m3_tools.py  # M3 full dApp tests
│   │   └── benchmark.py      # Evaluation framework
│   └── test_retrieval.py # Retrieval quality tests
├── docs/
│   └── mcp_tools_spec.md # MCP tools specification
├── apps/web/               # Hosted service (Cloudflare Workers + Next.js)
│   ├── src/lib/
│   │   ├── scraper.ts      # Web doc scraping (HTMLRewriter)
│   │   ├── github.ts       # GitHub repo scraping (Trees/Contents API)
│   │   ├── chunker.ts      # Document + code chunking
│   │   ├── ingestPipeline.ts # Ingestion orchestrator (sync + async queue paths)
│   │   └── vectorize.ts    # Search + embedding utilities
│   ├── src/app/api/admin/  # Admin APIs (sources, ingest, migrate)
│   ├── worker.ts           # Worker entry + cron + queue consumer handler
│   └── wrangler.prod.jsonc # Production config (D1, KV, Vectorize, Queue)
├── scripts/
│   ├── run_benchmarks.py     # Benchmark runner
│   ├── diff-migrate.ts       # Push chunks to CF Vectorize
│   ├── sync_sources.ts       # Sync sources.json to CF KV registry
│   └── ingest_m3_sources.py  # M3 source ingestion
├── data/
│   ├── raw/              # Raw scraped data (docs + curated repos)
│   ├── processed/        # Pre-processed chunks
│   └── chroma_db/        # ChromaDB vector store (generated locally, not in repo)
├── environment.yml       # Conda environment specification
├── pyproject.toml        # Project metadata and dependencies
└── .env                  # Environment variables (not committed)

Setup

1. Create Conda Environment

# Create and activate the environment
conda env create -f environment.yml
conda activate arbbuilder

Note: If you plan to refresh the knowledge base by scraping (optional), also install playwright:

playwright install chromium

2. Configure Environment Variables

Copy the example environment file and configure your API keys:

cp .env.example .env

Edit .env with your credentials:

OPENROUTER_API_KEY=your-api-key
NVIDIA_API_KEY=your-nvidia-api-key
DEFAULT_MODEL=deepseek/deepseek-v3.2
DEFAULT_EMBEDDING=baai/bge-m3
DEFAULT_CROSS_ENCODER=nvidia/llama-3.2-nv-rerankqa-1b-v2

3. Setup Data

The repository includes all data needed:

• Raw data (data/raw/): Documentation pages + curated GitHub repos
• Processed chunks (data/processed/): Chunks ready for embedding

Important: The ChromaDB vector database must be generated locally (it's not included in the repo due to binary compatibility issues across systems).

# Generate the vector database (required before using MCP tools)
python -m src.embeddings.vectordb

4. Verify MCP Server

Test that the MCP server starts correctly:

# Run the MCP server directly (press Ctrl+C to exit)
python -m src.mcp.server

You should see:

ARBuilder MCP Server started
Capabilities: 14 tools, 11 resources, 5 prompts

Optional: Refresh Data

If you want to re-scrape the latest documentation and code:

# Run full pipeline (web scraping + GitHub cloning)
python -m scraper.run

# Then preprocess the raw data
python -m src.preprocessing.processor

# And re-ingest into ChromaDB
python -m src.embeddings.vectordb --reset

Data Quality Filters: The pipeline applies a 3-layer filtering system to remove junk data (vendored crates, auto-generated TypeChain files, hex bytecode, lock files, and cross-repo duplicates). See docs/DATA_CURATION_POLICY.md for details.

Data Maintenance

Audit and clean up data sources:

# Audit: compare repos on disk vs config
python scripts/audit_data.py

# Show what orphan repos would be deleted
python scripts/audit_data.py --prune

# Actually delete orphan repos
python scripts/audit_data.py --prune --confirm

# Include ChromaDB stats in audit
python scripts/audit_data.py --chromadb

# GitHub scraper also supports audit/prune
python -m scraper.github_scraper --audit
python -m scraper.github_scraper --prune --dry-run

Fork & Migrate (SDK 0.10.0)

Fork community Stylus repos and migrate them to SDK 0.10.0:

# Dry run: show what would change without modifying anything
python scripts/fork_and_migrate.py --all --dry-run

# Migrate all 13 Stylus repos
python scripts/fork_and_migrate.py --all

# Migrate a specific repo
python scripts/fork_and_migrate.py --repo OffchainLabs/stylus-hello-world

# Re-verify already-forked repos after manual fixes
python scripts/fork_and_migrate.py --all --verify-only

Reports are saved to reports/fork_migration_*.json.

Quick Start (IDE Integration)

Option A: Self-Hosted (Full Control)

Run ARBuilder locally with your own API keys. No rate limits.

Step 1: Configure your IDE

Add the following to your MCP configuration file:

Cursor (~/.cursor/mcp.json):

{
  "mcpServers": {
    "arbbuilder": {
      "command": "/path/to/miniconda3/envs/arbbuilder/bin/python3",
      "args": ["-m", "src.mcp.server"],
      "env": {
        "OPENROUTER_API_KEY": "your-api-key",
        "PYTHONPATH":"/path/to/ArbBuilder"
      }
    }
  }
}

Claude Desktop (~/Library/Application Support/Claude/claude_desktop_config.json on macOS):

{
  "mcpServers": {
    "arbbuilder": {
      "command": "python",
      "args": ["-m", "src.mcp.server"],
      "cwd": "/path/to/ArbBuilder",
      "env": {
        "OPENROUTER_API_KEY": "your-api-key"
      }
    }
  }
}

Step 2: Restart your IDE

After saving the configuration, restart Cursor or Claude Desktop. The ARBuilder tools will be available to the AI assistant.

Step 3: Start building!

Ask your AI assistant:

• "Generate an ERC20 token contract in Stylus"
• "How do I deploy a contract to Arbitrum Sepolia?"
• "Write tests for my counter contract"

Option B: Hosted Service (Zero Setup)

Use our hosted API - no local setup required. Available at arbuilder.app.

1. Sign up at https://arbuilder.app and get your API key
2. Add to your MCP configuration:

{
  "mcpServers": {
    "arbbuilder": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://arbuilder.app/mcp",
               "--header", "Authorization: Bearer YOUR_API_KEY"]
    }
  }
}

The hosted service includes:

• 100 API calls/day (free tier)
• No local setup or Python environment required
• Always up-to-date with latest Stylus SDK patterns

Usage

Data Ingestion

Hosted (Worker-native): The hosted service at arbuilder.app has a built-in ingestion pipeline that runs automatically via cron (every 6 hours). Sources can also be manually ingested via the admin UI at /admin.

The pipeline uses two paths based on source size:

• Sync path (docs and repos ≤30 files): scrape → chunk → embed → upsert in a single Worker invocation (~40 subrequests)
• Async path (repos >30 files): scrape → chunk → save to KV → enqueue to CF Queue. The queue consumer processes embed/upsert in batches of 10 chunks (~4 subrequests each), with continue messages for additional file batches and a finalize message to update source status. This stays within the 50 subrequest/invocation limit on the Free plan.

Local (Python pipeline): For self-hosted setups, run the full data collection pipeline:

conda activate arbbuilder

# Run full pipeline (web scraping + GitHub cloning)
python -m scraper.run

# Preprocess and push to CF Vectorize
python -m src.preprocessing.processor
AUTH_SECRET=xxx npx tsx scripts/diff-migrate.ts --full

Data Sources

All data sources are defined in sources.json — the single source of truth for both the local Python pipeline and the hosted CF Worker ingestion. The file contains 84 curated sources (53 documentation pages + 31 GitHub repos) across 4 milestones.

Versioned repos (with multiple SDK branches) use a versions array:

{
  "url": "https://github.com/ARBuilder-Forks/stylus-hello-world",
  "versions": [
    { "sdkVersion": "0.10.0", "branch": "main" },
    { "sdkVersion": "0.9.0", "branch": "v0.9.0" }
  ]
}

Sync to hosted service:

ARBBUILDER_ADMIN_SECRET=xxx npx tsx scripts/sync_sources.ts
ARBBUILDER_ADMIN_SECRET=xxx npx tsx scripts/sync_sources.ts --dry-run
ARBBUILDER_ADMIN_SECRET=xxx npx tsx scripts/sync_sources.ts --remove-stale

Stylus (M1) — 17 docs + 19 repos

• Official documentation: docs.arbitrum.io (7 pages + gas-metering)
• All Stylus repos sourced from ARBuilder-Forks for resilience against upstream deletions
• 6 forks with SDK 0.10.0 branches: hello-world, vending-machine, erc6909, fortune-generator, ethbuc2025-gyges, WalletNaming
• 7 forks at original SDK version (0.8.4–0.9.0) with separate branch per version
• Production codebases: OpenZeppelin rust-contracts-stylus, stylus-test-helpers, stylusport, stylus-provider

Curation Policy:

• No meta-lists (awesome-stylus) — causes outdated code ingestion
• No unverified community submissions
• All code repos must compile with stylus-sdk >= 0.8.0
• SDK version tracked per-repo in sources.json
• All Stylus repos forked to ARBuilder-Forks org with forkedFrom provenance tracking

Stylus SDK Version Support:

Multi-Version Strategy:

• Branch-per-version: Forked repos maintain separate Git branches per SDK version (e.g., main for 0.10.0, v0.9.0 for original)
• Branch-aware scraping: CF Worker ingests each branch as a separate source entry
• Version-aware generation: generate_stylus_code and ask_stylus accept target_version to produce code for any supported SDK version
• Version-aware retrieval: Vector search boosts chunks matching the requested SDK version

Arbitrum SDK (M2) — 6 docs + 5 repos

• arbitrum-sdk, arbitrum-tutorials
• 3 community repos: arbitrum-api, orbit-bridging, cross-messaging
• Official bridging and messaging documentation (6 pages)

Full dApp Builder (M3) — 30 docs + 11 repos

• Backend: NestJS (5 docs), Express (3 docs), nestjs/nest, arbitrum-token-bridge
• Frontend: wagmi (5 docs), viem (4 docs), RainbowKit (4 docs), DaisyUI (5 docs) + 5 repos
• Indexer: The Graph (5 docs), graph-tooling, messari/subgraphs
• Oracle: Chainlink (4 docs), smart-contract-examples, chainlink

Orbit SDK (M4) — planned

API Access

Public MCP Endpoint (Free)

The MCP endpoint at /mcp is free to use and designed for IDE integration:

https://arbuilder.app/mcp

• Requires arb_ API key from dashboard
• Usage tracked per API key
• Rate limited per free tier (100 calls/day)

Transparency Page

View all ingested sources and code templates at arbuilder.app/transparency.

This public page provides:

• Ingested Sources: All documentation and GitHub repos in the knowledge base
• Code Templates: Verified Stylus templates with full source code
• Statistics: Chunk counts, SDK versions, and category breakdowns

Public API endpoints (no authentication required):

• GET /api/public/sources - List all active sources
• GET /api/public/templates - List all code templates
• GET /api/public/templates?code=true - Templates with full source code

Internal Direct API (Testing Only)

Direct API routes at /api/v1/tools/* are for internal testing only:

• Requires AUTH_SECRET in Authorization header
• Not for public use
• Used by CI/CD and internal validation scripts

MCP Capabilities

ARBuilder exposes a full MCP server with 13 tools, 11 resources, and 5 prompts for Cursor/VS Code integration.

Tools

M1: Stylus Development (6 tools)

M2: Arbitrum SDK - Bridging & Messaging (3 tools)

M3: Full dApp Builder (5 tools)

Example: Get Build/Deploy Workflow

{
  "workflow_type": "deploy",
  "network": "arbitrum_sepolia",
  "include_troubleshooting": true
}

Returns step-by-step commands:

# Check balance
cast balance YOUR_ADDRESS --rpc-url https://sepolia-rollup.arbitrum.io/rpc

# Deploy contract
cargo stylus deploy --private-key-path=./key.txt --endpoint=https://sepolia-rollup.arbitrum.io/rpc

Resources (Knowledge Injection)

MCP Resources provide static knowledge that AI IDEs can load automatically:

M1: Stylus Resources

M2: Arbitrum SDK Resources

M3: Full dApp Builder Resources

Prompts (Workflow Templates)

MCP Prompts provide reusable templates for common workflows:

How It Works

User: "Deploy my contract to Arbitrum Sepolia"
    ↓
AI IDE calls get_workflow(workflow_type="deploy", network="arbitrum_sepolia")
    ↓
Returns structured commands + troubleshooting
    ↓
AI IDE presents commands to user (user executes locally)

The MCP server provides knowledge about commands, not command execution. This ensures:

• User controls what runs on their machine
• No security risks from remote execution
• AI IDE knows exact commands without hardcoding

See docs/mcp_tools_spec.md for full specification.

User Guide

Generating Stylus Contracts

ARBuilder uses template-based code generation to ensure generated code compiles correctly. Instead of generating from scratch, it customizes verified working templates from official Stylus examples.

Available Templates:

Stylus SDK Version Support:

Pass target_version to tools for version-specific output:

User: "Generate a counter contract for SDK 0.9.0"
AI uses: generate_stylus_code(prompt="...", target_version="0.9.0")
Returns: Code using msg::sender(), .getter(), print_abi() patterns

Ask your AI assistant to generate contracts:

User: "Create an ERC20 token called MyToken with 1 million supply"

AI uses: generate_stylus_code tool
Returns: Complete Rust contract based on SimpleERC20 template with proper imports, storage, and methods

Getting Context and Examples

Search the knowledge base for documentation and code examples:

User: "Show me how to implement a mapping in Stylus"

AI uses: get_stylus_context tool
Returns: Relevant documentation and code snippets from official examples

Q&A and Debugging

Ask questions about Stylus development:

User: "Why am I getting 'storage not initialized' error?"

AI uses: ask_stylus tool
Returns: Explanation with solution based on documentation context

Generating Tests

Create test suites for your contracts:

User: "Write unit tests for this counter contract: [paste code]"

AI uses: generate_tests tool
Returns: Comprehensive test module with edge cases

Build/Deploy Workflows

Get step-by-step deployment guidance:

User: "How do I deploy to Arbitrum Sepolia?"

AI uses: get_workflow tool
Returns: Commands for checking balance, deploying, and verifying

Milestones

M2: Arbitrum SDK Integration

Cross-chain bridging and messaging support:

• ETH Bridging: L1 <-> L2 deposits and withdrawals
• ERC20 Bridging: Token bridging with gateway approvals
• L1 -> L3 Bridging: Direct L1 to Orbit chain bridging via double retryables
• Cross-chain Messaging: L1 -> L2 retryable tickets, L2 -> L1 messages via ArbSys
• Status Tracking: Message status monitoring and withdrawal claiming

# Example: Generate ETH deposit code
echo '{"method": "tools/call", "id": 1, "params": {"name": "generate_bridge_code", "arguments": {"bridge_type": "eth_deposit", "amount": "0.5"}}}' | python -m src.mcp.server

M3: Full dApp Builder

Complete dApp scaffolding with all components:

• Backend Generation: NestJS or Express with viem/wagmi integration
• Frontend Generation: Next.js 14 + wagmi v2 + RainbowKit v2 + DaisyUI
• Indexer Generation: The Graph subgraphs (ERC20, ERC721, DeFi, custom events)
• Oracle Integration: Chainlink Price Feeds, VRF, Automation, Functions
• Full Orchestration: Scaffold complete dApps with monorepo structure
• ABI Auto-Extraction: Contract ABI is parsed from Stylus Rust code and injected into backend/frontend
• ABI-Aware Generation: Indexer schema/mappings, frontend hooks, and backend routes are generated from contract ABI
• Compiler Verification: Docker-based cargo check loop catches and auto-fixes compilation errors
• Executable Scripts: Generated setup.sh, deploy.sh, and start.sh for one-command workflows
• CLI Scaffolding: setup.sh uses a scaffold-first, backfill pattern with official CLI tools (cargo stylus new, create-next-app, @nestjs/cli) to fill in config files our templates don't generate, with graceful fallback if tools aren't installed
• Env Standardization: Centralized env var config (PORT 3001, CORS, BACKEND_URL) across all components

Backend Templates:

• NestJS + Stylus contract integration
• Express + Stylus (lightweight)
• NestJS + GraphQL (for subgraph querying)
• API Gateway (cross-chain proxy)

Frontend Templates:

• Next.js + wagmi + RainbowKit base
• DaisyUI component library
• Contract Dashboard (admin panel)
• Token Interface (ERC20/721 UI)

Indexer Templates:

• ERC20 Subgraph (transfers, balances)
• ERC721 Subgraph (ownership, metadata)
• DeFi Subgraph (swaps, liquidity)
• Custom Events Subgraph

Oracle Templates:

• Chainlink Price Feed
• Chainlink VRF (randomness)
• Chainlink Automation (keepers)
• Chainlink Functions

# Example: Generate full dApp scaffold
echo '{"method": "tools/call", "params": {"name": "orchestrate_dapp", "arguments": {"prompt": "Create a token staking dApp", "components": ["contract", "backend", "frontend", "indexer"]}}}' | python -m src.mcp.server

# Example: Generate backend only
echo '{"method": "tools/call", "params": {"name": "generate_backend", "arguments": {"prompt": "Create a staking API", "framework": "nestjs"}}}' | python -m src.mcp.server

# Example: Generate frontend with contract ABI
echo '{"method": "tools/call", "params": {"name": "generate_frontend", "arguments": {"prompt": "Create token dashboard", "contract_abi": "[...]"}}}' | python -m src.mcp.server

Development

Running Tests

# Run all unit tests
pytest tests/ -m "not integration"

# Run retrieval quality tests
pytest tests/test_retrieval.py -v

# Run MCP tool tests (requires tool implementations)
pytest tests/mcp_tools/ -v

# Run template selection and validation tests
pytest tests/test_templates.py -v -m "not integration"

# Run template compilation tests (requires Rust toolchain + cargo-stylus)
pytest tests/test_templates.py -v -m integration

Template compilation tests require:

• Rust toolchain 1.87.0: rustup install 1.87.0
• WASM target: rustup target add wasm32-unknown-unknown --toolchain 1.87.0
• cargo-stylus: cargo install --locked cargo-stylus

Running Benchmarks

# Run all benchmarks
python scripts/run_benchmarks.py

# Run only P0 (critical) tests
python scripts/run_benchmarks.py --priority P0

# Run benchmarks for a specific tool
python scripts/run_benchmarks.py --tool get_stylus_context

Benchmark reports are saved to benchmark_results/.

Code Formatting

black .
ruff check .

Troubleshooting

Embedding Generation Errors

If you encounter errors like Error generating embeddings: RetryError or KeyError during vector database ingestion:

1. Check OpenRouter API Key

# Verify your .env file has a valid API key
cat .env | grep OPENROUTER_API_KEY

Ensure:

• The API key is correctly set (no extra spaces or quotes)
• Your OpenRouter account has credits
• The embedding model baai/bge-m3 is available on OpenRouter

2. Rate Limiting Issues

If you see HTTPStatusError with status 429, you're being rate limited. Solutions:

# Run with smaller batch size
python -m src.embeddings.vectordb --batch-size 25

# Or modify max_workers in vectordb.py to 1 for sequential processing

3. Enable Debug Logging

Add this to your script or at the start of your session to see detailed logs:

import logging
logging.basicConfig(level=logging.INFO)
# For more verbose output:
# logging.basicConfig(level=logging.DEBUG)

Scraper Errors

"Execution context was destroyed" errors

This is a browser navigation issue during scraping. The scraper will automatically retry. If it persists:

• The page may have heavy JavaScript that interferes with scraping
• These pages are skipped after retries; the scraper continues with other URLs

Git clone failures

If repository cloning fails:

# Check your network connection
ping github.com

# Try cloning manually to diagnose
git clone --depth 1 https://github.com/OffchainLabs/stylus-hello-world

# If behind a proxy, configure git
git config --global http.proxy http://proxy:port

Timeout errors

For slow connections, increase timeouts in the scraper config or reduce concurrent requests:

python -m scraper.run --max-concurrent 1

ChromaDB Issues

"Collection is empty" error

If you see collection is empty when using get_stylus_context tool:

# The vector database must be generated locally (it's not included in the repo)
# Run this command to populate the database:
python -m src.embeddings.vectordb

# If that doesn't work, try resetting first:
python -m src.embeddings.vectordb --reset

Import errors with opentelemetry

If you see TypeError: 'NoneType' object is not subscriptable when importing chromadb:

# This is usually a conda environment issue
# Make sure you're in the correct environment
conda activate arbbuilder

# Or reinstall chromadb
pip uninstall chromadb
pip install chromadb

Database corruption

If the vector database seems corrupted:

# Reset and re-ingest
python -m src.embeddings.vectordb --reset

CI/CD Workflows

maintenance.yml Jobs

License

MIT License - see LICENSE for details.

References

• Arbitrum Documentation
• Stylus Documentation
• ICP Coder - Reference implementation
• Stacks Builder - Reference implementation