Skip to main content

Guides · Guide

Your AI Side Project Doesn't Need a $500 Cloud Bill: The Cheap Infra Stack That Actually Works

Most AI apps are a frontend, a little backend and some model calls. You can run all of it for free or a few dollars a month with Vercel, Cloudflare, a small VPS, caching and local models. Here is the stack, and the agent skills that set it up for you.

By Skills-Hub Team4 min read
Guidesframeworks-on-cloudflareworkers-development
a tiny cozy server rack made of cardboard and fairy lights on a desk next to a piggy bank, a small glowing robot plugging in a cable, warm evening light

We earn commissions when you shop through the links below.

There's a specific kind of dread that comes from opening a cloud bill for a side project nobody uses yet. A managed database you forgot about, a GPU instance left running over the weekend, an API key that got hammered by a bot. It happens to everyone once.

The good news: almost every AI side project is the same shape. A frontend, a thin backend, a database, and some calls to a model. Each of those pieces has a free or nearly free option in 2026 if you pick deliberately. Here's the stack I'd start with, what each layer costs you, and the agent skills that let Claude Code set it up so you don't spend a weekend reading docs.

Frontend: free tiers are genuinely enough

Vercel and Cloudflare Pages both host a Next.js, Astro or plain React frontend on a free plan that comfortably covers a side project's traffic. Push to GitHub, get a URL, done. You only start paying when you have real usage, which is exactly when you want to be paying.

If you want your agent to handle the deploy, cloudflare-frameworks-on-cloudflare builds and deploys framework apps to Cloudflare, and vercel-agent-browser-vercel lets an agent open and test your Vercel deployments, even behind deployment protection.

  • Install: npx @skills-hub-ai/cli install cloudflare-frameworks-on-cloudflare
  • Install: npx @skills-hub-ai/cli install vercel-agent-browser-vercel

Backend: serverless until it isn't

a small friendly robot juggling lightweight glowing cubes labeled with no text, representing tiny serverless functions, floating over a laptop, clean flat illu…

For a thin API layer, serverless functions are hard to beat on price: Cloudflare Workers and Vercel Functions both include a free allowance, and you pay per request after that. That model is perfect while traffic is spiky and small. cloudflare-workers-development covers running and debugging Workers locally, which is where most of the time goes.

Serverless gets awkward when you need long-running jobs, background queues, websockets or a model server that stays warm. That's the point where a small always-on server becomes the cheaper and simpler option.

  • Install: npx @skills-hub-ai/cli install cloudflare-workers-development
  • Rule of thumb: serverless for request/response, a VPS for anything that runs for minutes

The always-on box: a small VPS

A single small virtual server runs your worker queue, a Postgres instance, a vector store and a reverse proxy at a flat monthly price that doesn't jump when traffic does. DigitalOcean Droplets are the easy on-ramp here: predictable monthly pricing, a clean dashboard, and one-click images for Docker. Hetzner is popular if you're comfortable with a more bare-bones experience.

Put everything in Docker Compose so the whole box can be rebuilt from one file. multi-language-standards-common-devops-operations gives your agent a provider-neutral playbook for exactly this kind of setup, across AWS, GCP, Cloudflare and Vercel.

  • Install: npx @skills-hub-ai/cli install multi-language-standards-common-devops-operations

The model bill is where the money actually goes

a robot carefully pouring glowing coins back into a jar from a funnel labeled with no text, a cache drawer full of saved answers beside it, warm flat illustrat…

Hosting is usually the small part. The real cost is tokens. Three habits cut it dramatically: cache identical and near-identical requests so you never pay twice for the same answer, send easy tasks to a smaller, cheaper model, and use provider-side prompt caching for long shared prompts.

devops-security-agent-skills-llm-caching implements multi-layer caching (exact match, semantic similarity and provider-side), and devops-security-agent-skills-llm-cost-optimization covers model selection, batching and routing. Either one usually pays for the hour it takes to add.

  • Install: npx @skills-hub-ai/cli install devops-security-agent-skills-llm-caching
  • Install: npx @skills-hub-ai/cli install devops-security-agent-skills-llm-cost-optimization

Or skip the API: run models locally

For internal tools, batch jobs and anything private, an open-weight model on your own machine costs nothing per token. devops-security-agent-skills-ollama-stack sets up Ollama with a web UI and GPU-aware tuning, and buildwithclaude-localaipicks helps you find which open models actually run well on your hardware before you download 40 gigabytes of the wrong one.

  • Install: npx @skills-hub-ai/cli install devops-security-agent-skills-ollama-stack
  • Install: npx @skills-hub-ai/cli install buildwithclaude-localaipicks

The starter stack

Frontend on Vercel or Cloudflare Pages, free. Thin API on Workers or Vercel Functions, free until it's busy. One small VPS for the database, queue and anything long-running. Caching and a cheaper default model in front of every LLM call. Local models for anything private or bulk. Set a billing alert on day one, and you'll never open that dreaded bill again.

Skills to try

  • frameworks-on-cloudflare: Build, migrate, and deploy web-framework applications on Cloudflare Workers. Use when a framework project such as Astro, SvelteKit, React Router, TanStack Star…
  • workers-development: Run, inspect, and debug Cloudflare Workers during local development with cf dev, wrangler dev, or the Cloudflare Vite plugin, using Local Explorer to read and …
  • vercel: Access and test Vercel deployments with agent-browser, behind Deployment Protection or at a different feature flag variant. Use when a preview or production UR…
  • llm-caching: Implement multi-layer LLM caching with exact match, semantic similarity, and provider-side prompt caching. Reduce API costs by 30–70%, cut latency, and improve…
  • llm-cost-optimization: Reduce LLM API and infrastructure costs through model selection, prompt caching, batching, caching, quantization, and self-hosting strategies. Track spend by t…
  • ollama-stack: Run local LLM workloads with Ollama, Open WebUI, and GPU-aware tuning for private development environments. Use when setting up private inference, local AI dev…

Sources

Written by

Skills-Hub Team

Skills-Hub is the open registry for AI coding skills, with SKILL.md files synced daily from Anthropic, Google, Microsoft, and 90+ official sources. Free + MIT.

Continue reading