The recommended stack
NOAN is the brain: facts, tasks, contacts, and the API your agents run on. These seven platforms are the hands we recommend pairing it with, and the exact stack our own agent fleet runs on. All of them have generous free tiers to start.
The anatomy
What an agent actually needs
Strip any agent down and there are only a few core elements: intelligence to reason with, a brain of verified facts to ground it, a database to remember state between runs, a channel to communicate on, and a voice if it talks. None of it is infrastructure you install. Every element is a hosted platform, and all you actually take from each one is an API key.
Intelligence
An AI model to reason, write, and decide. Frontier for judgment calls, small and fast for classification.
A brain
Verified facts to ground every output, and a task board that controls what the agent may act on.
A database
Working memory between runs: what was processed, who was contacted, when to cool down.
A way to communicate
Email, Slack, or your website. An agent that cannot speak to anyone is a cron job with opinions.
A voice
OptionalIf it talks to people out loud, add speech to text and text to speech. One more API, not a new architecture.
Somewhere to run
A schedule for routine work, an always-on process if it listens live. Never a server you manage.
# everything the agent needs
NOAN_API_KEY=nk_... # the brain: facts + tasks
ANTHROPIC_API_KEY=sk-... # the intelligence
SUPABASE_URL=https://... # working memory
SUPABASE_KEY=eyJ...
RESEND_API_KEY=re_... # the mail line
FIRECRAWL_API_KEY=fc-... # the researcher
# that's it. no servers, no SDKs required.
# each key is one sign-up on a free tier.That is the whole point of this stack: an agent is a small script plus a handful of environment variables. Sign up, create a key, paste it into the env file, and the platform is wired in.
Anthropic's coding agent. You describe what you want in plain language and it writes, runs, and tests the code in front of you.
This is how agents get built on NOAN. Point it at the API docs and the NOAN skill, hand it a build plan, and it writes the agent, wires the guardrails, and tests each part against your workspace.
Every agent in our fleet was built this way. Our setup guide takes you from a blank machine to a connected workspace.
Where the code lives, and with GitHub Actions, where it runs: scheduled jobs and event-triggered workflows on infrastructure you never manage.
Agents on NOAN are mostly cron scripts. Actions gives every agent a free, reliable runtime: a schedule fans out the fleet, an event dispatch picks up urgent work in seconds, and every run leaves a log.
Our entire fleet of eighteen agents runs on GitHub Actions. No servers, no container platform, just workflows in a repo.
Developer-first email: send from your own domain with an API call, and receive inbound mail as webhooks your code can read.
The moment an agent needs to email a customer, you need real deliverability and a reply path. Resend gives agents both: outbound from your domain, and Inbound so a reply becomes structured data the agent can act on.
Our onboarding, support, and sales agents all send through Resend, and a reply loop reads inbound mail so Verity answers or escalates within minutes.
Turns any website into clean, LLM-ready markdown: crawl a domain, scrape a page, and get structured content instead of raw HTML.
Facts cover your business; Firecrawl covers everyone else's. When an agent needs to understand a prospect, a competitor, or a customer's site, one call returns readable pages a model can reason over.
Our sales agents crawl a prospect's site before writing a single line of outreach, so every email is grounded in what that company actually says about itself.
A hosted Postgres database with auth, storage, and APIs on top. Relational state with almost zero setup.
NOAN holds your truth; Supabase holds your plumbing. Queues, cooldown timers, processed-message logs, anything an agent fleet needs to remember between runs that is state rather than knowledge.
Our agents coordinate through Supabase tables: dedupe state, relay queues, and usage logs, while every fact they reason from stays in NOAN.
Hosting for websites and web apps: push to a repo and the site builds, deploys, and scales on a global CDN.
Connect your site to your facts over the NOAN API and it becomes a reader of your brain: change a fact once and the live site updates with no redeploy. Netlify makes that site effortless to run.
The page you are reading is on Netlify, rendering live from the same facts our agents read. Merging to main is the whole deploy process.
Cloud hosting for long-running services: web servers, workers, and websocket apps that need to stay up, not just run on a schedule.
Most agents are fine as cron jobs, but some need a persistent process: a voice interface, a live chat brain, a webhook receiver. Render runs those as simple services with none of the platform overhead.
Our voice surfaces run on Render as always-on services, streaming speech in real time while reading the same NOAN fact base as the rest of the fleet.
How it fits together
One brain, boring hands
The pattern behind every agent we run: NOAN holds the verified facts and the task board, and each platform above does one job it is best at. Claude Code builds the agent, GitHub runs it, Resend lets it speak, Firecrawl lets it read the outside world, Supabase remembers its state, and Netlify and Render put your facts on the web. Every piece is replaceable because the truth never lives in any of them. It lives in NOAN.
You do not need all seven on day one. Start with Claude Code and a NOAN key, and add a platform when an agent hits a job NOAN alone does not cover. The Agent Planner recommends the right pieces for your specific idea.
Plan your first agent on this stack
Describe the agent you want and get a build plan that picks from this stack automatically: the facts to seed, the task trigger, and only the platforms your idea genuinely needs.