Daily recap for May 21, 2026

New tutorials

OpenRouter MCP Gateway for SMB Multi‑Tool Agents

Small businesses that want their AI agents to work with tools like CRMs and calendars often face a tangle of per-tool auth and security. This tutorial walks you through deploying a pre-built, multi-tenant MCP gateway that exposes any tool behind a unified authentication layer, with rate limiting and tenant isolation. It uses REAA's mcp-gateway-core and multi-tenant MCP middleware, instrumented with real-time metrics.

Under the hood: OpenRouter, Fastify, CLI — 209 tests, 96.25% coverage.

Read the tutorial → Download the code (zip)

xAI Grok Voice Agent for After-Hours Customer Support

Missing after-hours calls costs small businesses leads. This tutorial shows you how to deploy an AI receptionist powered by xAI Grok that answers calls, qualifies leads, and routes urgent issues to human staff via SMS—without 24/7 staffing. It uses LiveKit for real-time media, Deepgram for speech, Cartesia for speech synthesis, and REAA's confidence router, agent handoff, and budget engine to keep costs in check.

Under the hood: xAI Grok, Express — 99 tests, 100% coverage.

Read the tutorial → Download the code (zip)

Azure AI Spend Control for Multi-Model SMB Workflows

Running multiple Azure AI models can lead to unpredictable costs. This tutorial builds a budget-controlled chat server that enforces real-time per-scope spending limits, auto-downgrades to cheaper models when budgets tighten, and exports cost telemetry to Langfuse. It wraps Azure AI requests with REAA's agent-budget-engine and middleware, all in a Next.js app.

Under the hood: Azure AI, Express, Next.js — 79 tests, 98.15% coverage.

Read the tutorial → Download the code (zip)

vLLM Agent Eval Harness for Fine-Tuned Model Quality

Fine-tuning open models locally is only half the battle—you still need to verify quality before production. This CLI harness runs automated evaluations on models hosted behind vLLM, using GPT-4 as a judge to score against golden datasets, track token costs, and enforce regression gates. Built with REAA's agent-eval-harness-cli, judge, and cost modules.

Under the hood: vLLM, CLI — 126 tests, 100% coverage.

Read the tutorial → Download the code (zip)

Agnostic AI Cost Control for QuickBooks Online SMBs

If you're using multiple LLM providers, getting a single view of AI spend—and syncing it into QuickBooks—is manual work. This tutorial sets up an agnostic cost control service that monitors and caps spending across any provider, then posts categorized costs into QuickBooks Online as operational expenses. It uses REAA's budget engine, spend tracker, and QuickBooks sync adapter.

Under the hood: Agnostic, Express — 93 tests, 97.43% coverage.

Read the tutorial → Download the code (zip)

Agent Replay Toolkit for Deterministic Debugging

Record, replay, and debug agent runs with full traceability and time-travel debugging. This tutorial gives you a toolkit for capturing agent decisions deterministically, stepping through any run to find bugs fast. Built with TypeScript, ESM, Vitest.

Read the tutorial →

Voice Agent Pipeline for Real-Time Conversational AI

A modular toolkit for building, testing, and deploying voice agents with TTS, STT, and telephony integration. This tutorial walks you through creating a reusable pipeline for real-time conversational AI, with Vitest for testing. Stack: TypeScript, ESM, Vitest.

Read the tutorial →

New repos

sl-skills

sl-skills packages give AI coding agents a catalog of reusable skills—structured markdown files that tell the agent what a specific @reaatech package does, when to use it, and how to install it. Following Anthropic's progressive-disclosure pattern, the agent reads a one-line description to decide relevance, loading full details only when needed. Adopt this to stop writing ad-hoc tool-use instructions.

Browse the code on GitHub · View the catalog

Building blocks shipped

Media Pipeline MCP

Today's big drop is media-pipeline-mcp, a full-stack TypeScript toolkit for orchestrating AI media operations—image generation, audio synthesis, document extraction, 3D mesh generation—with provider routing, cost tracking, quality gates, and multi-tenancy. The 28 packages ship at v0.3.0 and include: the core framework (pipeline execution, budget enforcement, artifact registry), a server exposing 35+ media ops as MCP tools, and providers for OpenAI, Anthropic, ComfyUI (local GPU), Stability, ElevenLabs, Deepgram, Google, Fal, Replicate, Luma, Meshy, and Ollama. Supporting packages handle resilience (circuit breakers), security (auth/RBAC/rate limiting), cost ledgers, key vaults, observability, persistence (in-memory and Redis), provenance signing (C2PA), and video-gen with ffmpeg. It's a foundation for building your own AI media pipelines without stitching together SDKs.

Browse the media-pipeline-mcp catalog →

Browse the full catalog at reaatech.com/products.