Weekly recap, May 18, 2026 – May 24, 2026

This week we shipped 40 step-by-step tutorials for small-business AI — covering voice agents, multi-agent handoffs, document pipelines, cost controls, and more — plus 95 new building-block packages under 11 repos. If you're running operations, support, or accounting, there's something here you can try today.

New tutorials

Anthropic RAG Pipeline for SharePoint Knowledge Search

Help teams find answers buried in SharePoint libraries. This tutorial sets up a hybrid RAG system combining BM25 keyword search and vector similarity, with Claude synthesizing responses. A confidence router escalates low-confidence replies to Slack, while budget guards limit API costs. Ideal for SMBs drowning in document chaos.

Read the tutorial → Anthropic RAG Pipeline for SharePoint Knowledge Search · Download the code (zip)
Built with Anthropic, Next.js, 119 tests, 100% coverage.

Google Gemini Voice Agent for Twilio Call Handling

Replace after-hours voicemail with a Gemini-powered voice agent that answers calls, books appointments, and looks up FAQs — all through natural speech. Deepgram and ElevenLabs handle the audio, while REAA packages manage memory, confidence routing, and circuit breakers. A cost-effective 24/7 receptionist for small businesses.

Read the tutorial → Google Gemini Voice Agent for Twilio Call Handling · Download the code (zip)
Built with Google Gemini, Express, 91 tests, 98.95% coverage.

Azure AI Multi-Agent Handoff for Linear Issue Triage

Automate bug triage in Linear. A webhook triggers Azure OpenAI classifiers and specialized agents (debug, docs, scheduling) that auto-label and prioritize issues. Circuit breakers and budget caps keep the pipeline resilient, so engineers see ready-to-work tickets without the manual drill.

Read the tutorial → Azure AI Multi-Agent Handoff for Linear Issue Triage · Download the code (zip)
Built with Azure AI, Next.js, 124 tests, 99.65% coverage.

AWS Bedrock Multi‑Agent Handoff for Slack Support Triage

Route Slack messages to billing, technical, or account management specialists automatically. Using AWS Bedrock and a keyword-based confidence router, this Express app classifies and responds to support requests, handing off to human agents only when necessary.

Read the tutorial → AWS Bedrock Multi‑Agent Handoff for Slack Support Triage · Download the code (zip)
Built with AWS Bedrock, Express, 72 tests, 96.66% coverage.

Google Gemini AI Spend Control for SMBs

Add real-time budget enforcement to any Gemini-powered app. Pre-flight checks block or downgrade models when costs creep, while a live dashboard shows per-session spend via OpenTelemetry. Perfect for keeping unpredictable LLM costs in check.

Read the tutorial → Google Gemini AI Spend Control for SMBs · Download the code (zip)
Built with Google Gemini, Express, 79 tests, 96.55% coverage.

OpenAI Lead Intake with Adaptive Routing

Capture and qualify inbound leads from forms and attachments. OpenAI classifies intent, confidence routing sends hot leads straight to sales, and budget controls cap costs. Includes HubSpot sync and duplicate prevention — an end-to-end lead processing engine.

Read the tutorial → OpenAI Lead Intake with Adaptive Routing · Download the code (zip)
Built with OpenAI, Express, 199 tests, 100% coverage.

AWS Bedrock Code Sandbox for SMB Financial Report Generation

Let non-technical staff ask natural-language questions about financial data and get safe, formatted reports. Bedrock generates Python code executed in an E2B sandbox, with budget enforcement, caching, and structured output repair — no spreadsheet errors.

Read the tutorial → AWS Bedrock Code Sandbox for SMB Financial Report Generation · Download the code (zip)
Built with AWS Bedrock, Next.js, 85 tests, 98.52% coverage.

OpenAI Voice Agent for Twilio Customer Support Calls

Answer every inbound call with an OpenAI-powered agent that classifies intent, enforces per-call budgets, and falls back to voicemail on degradation. Twilio webhooks, Deepgram transcription, and Cartesia TTS create a production-ready voice pipeline for small support teams.

Read the tutorial → OpenAI Voice Agent for Twilio Customer Support Calls · Download the code (zip)
Built with OpenAI, Express, 89 tests, 98.45% coverage.

Perplexity Code Sandbox for SMB Data Reporting

Run AI‑generated analytics code safely. Perplexity creates Python or SQL, a Daytona sandbox executes it, and reliability wrappers from REAA (circuit breakers, budget controls, cache) ensure nothing blows up. A self-service reporting tool for analysts.

Read the tutorial → Perplexity Code Sandbox for SMB Data Reporting · Download the code (zip)
Built with Perplexity, Next.js, 128 tests, 94.81% coverage.

Cohere Reliability Suite for SMB Customer Support

Make Cohere-powered support agents survive outages. This reliability wrapper adds model-agnostic fallback to Azure OpenAI, circuit breakers, output repair, duplicate prevention, and daily budget caps — all orchestrated by Trigger.dev durable workflows.

Read the tutorial → Cohere Reliability Suite for SMB Customer Support · Download the code (zip)
Built with Cohere, Express & Next.js, 117 tests, 94.48% coverage.

OpenRouter MCP Gateway for SMB Multi‑Tool Agents

Let OpenRouter‑powered agents safely talk to your CRM, email, and calendar. This multi-tenant MCP gateway provides per-tenant auth, rate limiting, and tool visibility, while the CLI helps operators manage tenants. One server, many tools, no custom plumbing.

Read the tutorial → OpenRouter MCP Gateway for SMB Multi‑Tool Agents · Download the code (zip)
Built with OpenRouter, Fastify, 209 tests, 96.25% coverage.

xAI Grok Voice Agent for After-Hours Customer Support

Deploy an AI receptionist that answers calls, qualifies leads, and routes urgent issues via Twilio SMS. Grok handles conversation, LiveKit manages real-time media, and REAA packages enforce budgets, cache common responses, and secure the handoff to human staff.

Read the tutorial → xAI Grok Voice Agent for After-Hours Customer Support · Download the code (zip)
Built with xAI Grok, Express, 99 tests, 100% coverage.

Azure AI Spend Control for Multi-Model SMB Workflows

Prevent runaway Azure AI costs across models. This budget-controlled server checks every chat request, auto-downgrades to cheaper models when limits near, and exports cost telemetry to Langfuse. Includes a health server and admin spend summaries.

Read the tutorial → Azure AI Spend Control for Multi-Model SMB Workflows · Download the code (zip)
Built with Azure AI, Express, 79 tests, 98.15% coverage.

vLLM Agent Eval Harness for Fine-Tuned Model Quality

Run CI/CD-style evaluations on local fine-tuned LLMs behind vLLM. This CLI tool harnesses GPT-4 as a judge, tracks costs, enforces regression gates, and outputs Langfuse traces. Ensure every model change meets quality bars before production.

Read the tutorial → vLLM Agent Eval Harness for Fine-Tuned Model Quality · Download the code (zip)
Built with vLLM, CLI, 126 tests, 100% coverage.

Agnostic AI Cost Control for QuickBooks Online SMBs

Monitor and cap spending across any LLM provider, then sync categorized costs into QuickBooks Online. Budget engines enforce per-project limits, OpenTelemetry exports cost spans, and a daily cron pushes journal entries — giving SMBs financial control without spreadsheets.

Read the tutorial → Agnostic AI Cost Control for QuickBooks Online SMBs · Download the code (zip)
Built with agnostic providers, Express, 93 tests, 97.43% coverage.

AWS Bedrock Security Guardrails for Stripe Payment Agents

Intercept every Stripe API call from AI agents and check it against safety policies: PII redaction, prompt injection defense, and policy gates (no refund over $500 without human approval). An MCP gateway with circuit breakers and idempotency prevents costly errors.

Read the tutorial → AWS Bedrock Security Guardrails for Stripe Payment Agents · Download the code (zip)
Built with AWS Bedrock, Express, 95 tests, 96.78% coverage.

Azure AI Reliability Suite for SMB AI Operations

Proactive incident detection, self-healing, and cost-aware failure recovery for Azure AI agents. This suite layers circuit breakers, idempotency, and automated runbooks over Trigger.dev durable workflows, giving SMBs production reliability without a dedicated SRE team.

Read the tutorial → Azure AI Reliability Suite for SMB AI Operations · Download the code (zip)
Built with Azure AI, Express & Next.js, 102 tests, 97.70% coverage.

Vertex AI Document Pipeline for DocuSign SMB Contract Review

Automatically extract, summarize, and validate key clauses from incoming DocuSign contracts. Hybrid RAG indexes PDFs and Gemini extracts structured data; confidence scoring flags low-confidence results for attorney review. Stop manual contract review.

Read the tutorial → Vertex AI Document Pipeline for DocuSign SMB Contract Review · Download the code (zip)
Built with Vertex AI, Express, 126 tests, 98.86% coverage.

Anthropic Agent Mesh for E-commerce Fraud Detection

Three specialist Claude agents (anomaly detection, account takeover, chargeback risk) collaborate through the REAA agent mesh to evaluate orders in real time. A confidence gate automates approve/flag/block decisions, with human review escalation built in.

Read the tutorial → Anthropic Agent Mesh for E-commerce Fraud Detection · Download the code (zip)
Built with Anthropic, Next.js, 93 tests, 100% coverage.

Azure AI Code Sandbox for SMB Analytics

Secure code execution for business analytics. Azure OpenAI generates code, E2B runs it, and REAA budget/circuit-breaker packages enforce spending limits and isolate failures. A Next.js dashboard lets business users submit queries and track spending.

Read the tutorial → Azure AI Code Sandbox for SMB Analytics · Download the code (zip)
Built with Azure AI, Next.js, 65 tests, 93.63% coverage.

OpenAI Lead Intake Agent for SMB Real Estate

Automated lead capture for real estate agencies. OpenAI extracts structured lead data from forms and PDFs, classifies buyer/seller/renter intent, prevents duplicates with idempotency, and syncs to HubSpot. Drastically reduce manual data entry.

Read the tutorial → OpenAI Lead Intake Agent for SMB Real Estate · Download the code (zip)
Built with OpenAI, Next.js, 102 tests, 97.26% coverage.

AWS Bedrock Knowledge Agent for AppFolio Tenant Inquiries

A self-service AI chat for tenants that answers policy, maintenance, and lease questions using AppFolio data. Hybrid RAG indexes documents, Bedrock Claude answers with citations, and budget tracking keeps costs under control. Reduces property manager workload.

Read the tutorial → AWS Bedrock Knowledge Agent for AppFolio Tenant Inquiries · Download the code (zip)
Built with AWS Bedrock, Next.js & Express, 128 tests, 100% coverage.

Ollama Code Sandbox for SMB Financial Analysis

On-premises code execution for financial queries. A local Ollama LLM generates Python/SQL, a Daytona sandbox runs it, and budget/circuit-breaker packages enforce limits — all data stays on your hardware. Ideal for privacy-sensitive SMBs.

Read the tutorial → Ollama Code Sandbox for SMB Financial Analysis · Download the code (zip)
Built with Ollama, Express, 56 tests, 95.27% coverage.

OpenAI Invoice Extraction for Xero SMB Accounting

Turn PDF and image invoices into Xero entries automatically. GPT-5.2 vision parses invoice data, structured output repair fixes malformed JSON, confidence routing posts high-confidence lines, and budget controls cap spend. Eliminate manual data entry.

Read the tutorial → OpenAI Invoice Extraction for Xero SMB Accounting · Download the code (zip)
Built with OpenAI, Next.js & Express, 123 tests, 100% coverage.

Cohere Voice Agent for Shopify Order Status Checks

Customers call and ask "where's my order?" — Cohere's voice agent pulls live Shopify status, responds naturally, and escalates to Zendesk if needed. Circuit breakers, per-call budgets, and duplicate prevention keep it reliable and affordable.

Read the tutorial → Cohere Voice Agent for Shopify Order Status Checks · Download the code (zip)
Built with Cohere, Express, 111 tests, 100% coverage.

OpenAI Multi-Agent Handoff for E-commerce Order Handling

Automate order status, modifications, and cancellations with three specialist agents orchestrated by Mastra. OpenAI classifies intent, an LLM cache handles repeats, and idempotency ensures no double-refunds. Keeps support teams focused on exceptions.

Read the tutorial → OpenAI Multi-Agent Handoff for E-commerce Order Handling · Download the code (zip)
Built with OpenAI, Next.js & Express, 109 tests, 97.26% coverage.

Anthropic Document Pipeline for SMB Lease Abstraction

Extract key lease terms from PDFs automatically. Claude parses documents with retrieval-augmented context, low-confidence extractions are flagged for review, and budget/circuit breakers protect costs. Turns hours of manual review into seconds.

Read the tutorial → Anthropic Document Pipeline for SMB Lease Abstraction · Download the code (zip)
Built with Anthropic, Next.js & Express, 146 tests, 99.70% coverage.

Mistral AI Agent Reliability Suite for SMB Customer Support

Keep Mistral support agents running when downstream APIs fail. Circuit breakers isolate tool failures, idempotency prevents duplicate actions, and Trigger.dev jobs generate incident runbooks. A Next.js health endpoint exposes Kubernetes probes.

Read the tutorial → Mistral AI Agent Reliability Suite for SMB Customer Support · Download the code (zip)
Built with Mistral, Hono, 63 tests, 98.51% coverage.

Vertex AI Voice Agent for Twilio Field Service Scheduling

Automate field service booking over phone calls. LiveKit + Vertex AI power a real-time voice agent that schedules appointments in Google Calendar, routes to booking/rescheduling specialists, and enforces per-call budgets. A Next.js dashboard monitors everything.

Read the tutorial → Vertex AI Voice Agent for Twilio Field Service Scheduling · Download the code (zip)
Built with Vertex AI, Next.js & Express, 135 tests, 98.95% coverage.

Anthropic MCP Server for SMB ERP Tool Access

Expose HubSpot CRM as type-safe MCP tools that Claude can call. This server adds API-key auth, per-tenant tool allowlists, rate limiting, and audit logging — no custom integration code. Secure ERP access for AI agents in hours, not weeks.

Read the tutorial → Anthropic MCP Server for SMB ERP Tool Access · Download the code (zip)
Built with Anthropic, Hono, 90 tests, 95.50% coverage.

Vertex AI Reliability Suite for SMB Agent Operations

Wrap AI agents in production-grade reliability. Circuit breakers, idempotent retries, structured output repair, and automated incident playbooks — all on Vertex AI and orchestrated by Inngest. No SRE team required.

Read the tutorial → Vertex AI Reliability Suite for SMB Agent Operations · Download the code (zip)
Built with Vertex AI, Hono & Next.js, 71 tests, 98.38% coverage.

String Utils

Two small, well-tested packages of TypeScript string helpers — core pure functions and a tiny CLI wrapper. A minimal utility for when you need quick string operations without pulling in a large library.

Read the tutorial → String Utils

Structured Function Calling for AI Agents

Enforce schema contracts, validate tool outputs, and orchestrate function calls across LLMs. This monorepo gives you middleware to make function-calling reliable and type-safe, with ESM builds and full test coverage.

Read the tutorial → Structured Function Calling for AI Agents

Browser Automation Toolkit for AI Agents

Give your AI agent reliable browser interaction primitives: page navigation, form filling, content extraction, and human-in-the-loop supervision. Stealthy and resilient, with CLI and Node.js APIs.

Read the tutorial → Browser Automation Toolkit for AI Agents

REAA Structured Logger

A typed structured-log builder with field redaction, plus a Pino transport. Core logger and adapter in one monorepo, ideal for production log pipelines with sensitive data masking.

Read the tutorial → REAA Structured Logger · Download the code (zip)

REAA Rate Limiter

Pluggable rate limiting: token-bucket + sliding-window algorithms, with an ioredis backend. Core and adapter packages for adding rate limits to any Node.js service.

Read the tutorial → REAA Rate Limiter · Download the code (zip)

Code Execution Sandbox for AI Agents

Provide secure, isolated code execution for LLM-generated code. Pluggable backends (E2B, local, etc.) and MCP integration make it a building block for any agent that needs to run untrusted code.

Read the tutorial → Code Execution Sandbox for AI Agents

Agent Replay Toolkit for Deterministic Debugging

Record, replay, and debug agent runs with full traceability. Time-travel debugging for AI agents — capture every step and replay with different code to find the root cause.

Read the tutorial → Agent Replay Toolkit for Deterministic Debugging

Voice Agent Pipeline for Real-Time Conversational AI

A modular toolkit for building and testing voice agents with TTS, STT, and telephony integration. Everything you need to compose a voice agent from components, with CLI and test utilities.

Read the tutorial → Voice Agent Pipeline for Real-Time Conversational AI

RAG Pipeline Toolkit for Production AI

Vendor-agnostic, modular building blocks for retrieval-augmented generation. Pluggable embeddings, vector stores, and rerankers — pick your stack and compose a pipeline.

Read the tutorial → RAG Pipeline Toolkit for Production AI

New repos

sl-skills

Agent-consumable skills for the REAA Standard Library — auto-generated from the live catalog at reaatech.com. These structured markdown files tell your coding agent what each @reaatech package does, when to use it, and how to install it. Follows Anthropic's progressive-disclosure pattern: the agent reads a one-line description to decide relevance, then loads the full body only when it reaches for that skill, keeping context small and decisions fast.

Browse the code · Catalog page

Building blocks shipped

multi-tenant-mcp

A complete toolkit for building multi-tenant MCP servers: tenant resolution, rate limiting, tool visibility, cost accounting, observability, and middleware. 9 packages shipped this week, including the core middleware that enforces access control and rate limits per tenant, and the tenant resolver supporting JWT, API key, and header-based identification.

Browse the family →

otel-cost-exporter

Real-time LLM cost tracking via OpenTelemetry. 5 packages: core types, bundled pricing tables for all major providers, a cost calculator, a span processor/collector, and a CLI. Drop it into your OTel pipeline to get per-request cost dashboards.

Browse the family →

otel-genai-semconv

Instrument LLM providers (Anthropic, Bedrock, OpenAI, Vertex AI) with GenAI OpenTelemetry semantic conventions. 9 packages: shared types, SDK initialization, token counting, custom exporters for Arize Phoenix, Langfuse, and Google Cloud Trace — plus instrumentation wrappers.

Browse the family →

prompt-injection-bench

Standardized benchmarking for prompt injection defenses. 9 packages: attack corpora with obfuscation, defense adapters, scoring, leaderboards, parallel runners, and an MCP server. Compare your guardrails quantitatively.

Browse the family →

media-pipeline-mcp

A massive, modular media processing framework with 35+ operations across image generation/editing, audio TTS/STT, video generation, document extraction, and 3D mesh generation. 28 packages shipped: core orchestration, provider adapters (Anthropic, Google, OpenAI, fal, Stability, Deepgram, ElevenLabs, Replicate, ComfyUI, Ollama, Luma, Meshy), quality gates, cost tracking, provenance signing, and a full MCP server.

Browse the family →

prompt-version-control

Git-like version control for AI prompts. 5 packages: shared types/Schemas, a Hono API server with eval-gated promotion and A/B deployment, a TypeScript client, a CLI, and an MCP server. Manage prompt templates across draft, staging, and production with full audit trails.

Browse the family →

rag-eval-pack

Comprehensive RAG evaluation toolkit. 10 packages: heuristic metrics (faithfulness, relevance), LLM-as-judge with multi-model consensus, quality gates, cost tracking, dataset management, observability, and an MCP server — everything you need to quantify and gate RAG pipeline quality in CI.

Browse the family →

secret-rotation-kit

Zero-downtime secret rotation across AWS, GCP, and HashiCorp Vault. 7 packages: provider adapters, a core rotation manager with propagation verification, a sidecar REST API with SSE, and Prometheus metrics. Avoid manual key rotation and eliminate risks.

Browse the family →

session-continuity-kit

Manage multi-turn AI conversations with token budgets, context compression, and persistence. 6 packages: a session manager core, storage adapters for DynamoDB, Firestore, Redis, and in-memory (for dev), plus tokenizers for OpenAI and Anthropic. Keep agent context across calls without blowing context windows.

Browse the family →

structured-output-repair

Fix malformed JSON from LLMs against Zod schemas. Two packages: a core repair pipeline with multiple strategies, and an MCP server for easy integration. Guarantee your structured outputs match the contract.

Browse the family →

voice-agent-kit

Build real-time voice agents with STT, TTS, Twilio telephony, and MCP integration. 5 packages: core pipeline orchestration, Deepgram/AWS/Google STT providers, TTS providers, a Twilio media stream handler, and an MCP client. Plug-and-play components for your next voice bot.

Browse the family →

Browse the full catalog at reaatech.com/products.