AWS Builder Center
Building "Vercel for AWS": How I Let an AI Agent Provision Real AWS Infrastructure (And What It Taught Me About Least Privilege)

Building "Vercel for AWS": How I Let an AI Agent Provision Real AWS Infrastructure (And What It Taught Me About Least Privilege)

A weekend build where Amazon Bedrock reads a GitHub repo, decides what AWS services it needs, and provisions them live — plus the security mistake I almost shipped, and what AWS's own defaults taught me about IAM.

The problem I kept running into

AWS is, hands down, the most powerful cloud platform out there — and it's genuinely painful to deploy to if you're trying to do it properly. Before you've written a single line of infrastructure code, you already have to decide: S3? Lambda? ECS? Then you're configuring CloudFront by hand, writing IAM policies you half-copied from some five-year-old Stack Overflow answer (and which are almost certainly more permissive than you meant), and you don't find out what any of it costs until the bill lands in your inbox.
Now compare that to  git push  on Vercel.
It's not that AWS is missing anything — every primitive you need is right there. Nobody's just built the layer that makes choosing between them effortless. So for AWS's "First Commit" hackathon, I decided to build that layer myself. The project is called Anchor: you paste a GitHub repo, an AI agent figures out what it actually needs, and it deploys straight to real AWS infrastructure — no Terraform, no console clicking, no guessing.

How it actually works

The flow ended up being a lot simpler than I expected going in.
First, Anchor reads the repo — just the top-level file listing, plus  package.json 's scripts and dependencies if one exists. No cloning yet. Just enough signal to know what it's looking at.
That manifest then goes to Amazon Bedrock for classification. (I ended up landing on Amazon Nova through the Converse API — there's a story behind that switch, more on it below.) Bedrock hands back structured JSON: is this a static frontend? What's the build command? Where does the build output actually land once it's done?
From there it's build and provision. For static frontends, Anchor clones the repo, runs the build, spins up a dedicated private S3 bucket, uploads the output, and sits a CloudFront distribution with Origin Access Control in front of it. No public S3 buckets — not once, not even for a second.
And the whole time, every stage gets written to DynamoDB — not just a final "done" flag, but live status at each step. That's what powers the real-time progress UI on the frontend. It's not a spinner pretending something's happening. It's an actual checklist, and each checkmark reflects something that genuinely just finished in the backend.
GitHub repo → manifest → Bedrock classification → build → S3 (private) → CloudFront (OAC) → live URL ↓ DynamoDB (status + history)
Any opinions in this article are those of the individual author and may not reflect the opinions of AWS.
Enjoyed reading this content? Let the author know!

Your likes, comments, shares, and saves help creators reach more builders.

Loading recommendations

Loading article