Vibe Coding on AWS: From Kiro Prompt to Production in 7 Steps (2027)

Superblocks Team
+2

Multiple authors

November 18, 2026

0 min read

Copied
0:00

Vibe coding on AWS now centers on Kiro and Amazon Bedrock, since AWS is retiring Amazon Q Developer's IDE plugins. I read AWS's pricing pages, sunset notices, and docs to map the path from prompt to a production app that can pass your security review.

What is vibe coding on AWS?

Vibe coding on AWS means describing an app in plain language to an AI agent, then running what it generates on AWS infrastructure. In practice, the work splits into generating the code and getting it into a production account safely.

AWS handles code generation with Kiro, and Amazon Bedrock supplies the models your app calls once it runs. The production side is mostly yours: secrets, sign-in, hosting, infrastructure, monitoring, and a review gate before anything ships.

Part 1 explains which AWS tools to use now that the lineup has changed. Part 2 takes an app from prompt to production in 7 steps, then gets into the security review and who should approve AI-written code before release.

Part 1: The AWS tools for vibe coding

AWS changed its product lineup in 2026, so a 2025 tutorial can point you to a tool you can no longer sign up for. This is where each one stands as of November 2026:

Tool What it does Paid plans start at Status
Kiro Agentic IDE, CLI, and web agent $20 per user per month Open to new users
Amazon Q Developer IDE plugins, AWS Console assistant $19 per user per month Closed to new subscriptions
Amazon Bedrock Foundation models and agents for your app Pay per use Open to new users

Kiro and spec-driven development

Kiro is the tool AWS directs Q Developer users to. It comes as an IDE, a CLI, a web interface, a mobile app, and Kiro Crew. You can pick Anthropic Claude or OpenAI GPT models, plus open-weight options like DeepSeek and Qwen, or let Auto route each request.

Its default mode works like other chat-based coding agents. You prompt, it writes code, and a reviewer sees the plan once the code exists.

For production work, spec mode reverses that order and shows the reviewer the plan before any code is written. According to the Kiro specs docs, every spec produces three files:

  • requirements.md holds user stories with acceptance criteria (or bugfix.md for bug analysis).
  • design.md lays out the architecture, sequence diagrams, data flow, error handling, and the testing strategy.
  • tasks.md is a trackable implementation plan you can run one task at a time or all at once.

In the IDE, the requirements and design documents have a Review button that opens Markdown Review, where you comment on exact passages and send the comments back to Kiro. An Analyze Requirements check looks for contradictions and missing requirements before design starts.

That review step lets a tech lead reject a design.md that stores customer data in the wrong place before implementation starts. Kiro also offers Quick Spec, which generates all three files in one pass "without approval gates," so don't use it on work that touches production data.

Once a spec is approved, steering files in .kiro/steering/ (product.md, tech.md, structure.md) keep Kiro on your stack and conventions on every request. Custom agents only load them if you add them to the agent's resources configuration.

Hooks handle the checks that should run every time. Agent hooks in .kiro/hooks/ run commands on events like PostFileSave or PreToolUse, so a linter or security scanner starts without anyone having to launch it.

Pricing: Kiro Pro costs $20 per user per month, billed monthly, with 1,000 credits. Pro+ is $40 (2,000 credits), Pro Max is $100 (5,000), and Power is $200 (10,000), all per user per month. Kiro Free includes 50 credits per month.

Extra credits cost $0.04 each. Unused plan credits don't roll over, while purchased add-on credits carry over and expire 12 months after purchase.

Kiro Enterprise works differently. It uses custom pricing and is billed through AWS, and it requires sign-in through AWS IAM Identity Center or an external identity provider.

Amazon Q Developer after the IDE plugins

On April 30, 2026, AWS announced the end of support for Amazon Q Developer's IDE plugins and paid subscriptions, on this timeline:

  • May 15, 2026: New Free Tier accounts and new subscriptions blocked. Existing Pro subscriptions can still add seats.
  • April 30, 2027: IDE plugins and paid subscriptions reach end of support, along with Java upgrades and .NET porting through those plugins.

The CLI moved first. The Q Developer CLI has been rebranded to Kiro, including the CLI behind AWS's own 2025 vibe coding walkthrough, which generated a fact-checking app from a q chat prompt in about 12 minutes.

Q Developer in the AWS Management Console, the AWS Console Mobile App, and the Slack and Microsoft Teams chat apps keep running. So Q stays useful for questions like "why is this Lambda timing out," while Kiro takes over the coding.

Pricing: If your company already pays for it, Pro costs $19 per user per month, billed monthly, with the full month charged if a user cancels mid-month.

The plan also caps code transformation. Per that pricing page, Pro includes 4,000 lines of Java upgrade transformation per user per month, pooled at the account level. Superblocks' Amazon Q Developer pricing breakdown lists the rest of the usage caps.

Amazon Bedrock for the models inside your app

Kiro writes the code. Bedrock is what your app calls when it needs a model at runtime, whether that's summarizing a ticket or classifying an invoice.

AWS describes Bedrock as a platform for generative AI apps and agents with access to hundreds of foundation models. Bedrock AgentCore handles agent infrastructure, and Bedrock Guardrails can block up to 88% of harmful content, per AWS.

Pricing: Bedrock bills by usage, with no seat price. It's available on both the AWS Free plan and Paid plan, while AgentCore is Paid-plan only.

For the agent side, AWS also publishes vibe coding guidance with 6 MCP servers that plug into Kiro or other MCP-compatible agents. I'd install the AWS Pricing MCP Server first for cost estimates, then the AWS CDK MCP Server, which checks infrastructure code against CDK Nag rules.

What you need before starting

Part 2 assumes you already have a working app, made in Kiro or exported from another tool. Before you touch AWS, have these in place:

  • Separate dev and production AWS accounts, so an agent working in the sandbox has no access to production.
  • Single sign-on for your team's AWS access. Kiro Enterprise requires it, and SSO removes the need for long-lived access keys on laptops.
  • A Git repository, because each step below should end in a commit someone can review.
  • A named reviewer who signs off before release. Step 7 and the governance section explain who that should be.

Time required: In my estimate, a small internal app with one data store takes a few days of engineering time. Regulated data adds time, mostly in Steps 1 and 3.

Part 2: How to take a vibe-coded app to production on AWS in 7 steps

Each step depends on the one before it. The auth checks in Step 3 use the endpoint list from Step 1, and the infrastructure code in Step 5 describes the hosting you pick in Step 4.

Step 1: Audit what the AI wrote

Start with an inventory. Find the entry point (server.js, app.py, main.py, or a start script in package.json), then list the endpoints, data stores, third-party API calls, and dependencies.

For each endpoint, write down who can call it and what data it returns. This list drives Steps 2 and 3, and if the agent added an admin route no one requested, it shows up here.

While you have the code open, look for hardcoded localhost URLs, missing environment variables, and CORS errors, which Superblocks' guide to deploying vibe-coded apps flags as recurring bugs in AI-generated apps. Fix them now, before they surface as a failed deploy.

Pro tip: If the app came from Kiro's default mode, ask Kiro to write a design.md for the existing code and hand it to your reviewer as a reference document to compare against the code.

Step 2: Move secrets out of the code

Claude Code-assisted public commits leak secrets at about twice the baseline rate. GitGuardian found that 3.2% of them contained a secret, against 1.5% across all public commits.

To keep keys out of future commits, add .env to .gitignore before the first commit, then move every key, database password, and token into AWS Secrets Manager. It costs $0.40 per secret per month plus $0.05 per 10,000 API calls, and your code reads the value at runtime.

Plain environment variables leak too. In a Kiro-made events platform on AWS Builder Center, a code review found a captcha key stored as a plaintext Lambda environment variable, readable by anyone with basic read access.

Then rotate every key that ever appeared in a prompt, a commit, or a chat log, because a copy may already exist somewhere you don't control.

Step 3: Add proper sign-in and permissions

The first of the 7 risks in this vibe coding security guide is missing authorization, where generated endpoints confirm a session but never check what that user may access. Fixing it takes two layers, sign-in and permissions.

For sign-in, use Amazon Cognito or federate directly with your corporate identity provider through SAML or OIDC. Internal apps shouldn't need their own username and password table.

For permissions, enforce roles in one middleware layer on the server, then test every endpoint from Step 1 with no token and with a low-privilege token.

The app's own AWS access needs the same discipline. Give it an IAM role scoped to the exact resources it touches, and review each policy the agent writes. A wildcard policy removes permissions errors, which is why an agent that hits a permissions error may write one.

Step 4: Choose where it runs on AWS

Match the host to the app's architecture:

App shape AWS service Watch out for
Static site, SPA, or Next.js SSR app AWS Amplify Hosting Edge middleware and heavy backends
Full-stack app in a container Amazon ECS Express Mode Fargate and load balancer costs
Small APIs or event-driven jobs AWS Lambda Ephemeral file system

Amplify Hosting connects to a Git repository, creates a preview for each pull request, and serves the app from CloudFront on your own domain. For Next.js 12 through 15, it supports SSR and API routes, plus middleware, which is enough for an internal tool that doesn't depend on Edge middleware.

Containerized apps go to ECS Express Mode instead. It needs a container image plus two IAM roles, then creates a Fargate service, a load balancer with TLS, auto scaling, and monitoring.

Express Mode itself has no extra charge. You pay for the Fargate and load balancer resources it creates, which is the cost the table above flags.

Older tutorials still point to AWS App Runner, which AWS has closed to new customers. For new container apps, AWS recommends ECS Express Mode.

Next, move the data. Lambda and Fargate both use temporary storage, so if the app stores data in a local SQLite file, that file disappears when the instance does. Point the app at Amazon RDS or Aurora before launch.

Step 5: Define the infrastructure as code

Everything from Steps 2 to 4 goes into AWS CDK or Terraform, so the production setup is defined in a file someone reviewed. Pick whichever your platform team already uses; Kiro can draft either one.

Let Kiro write the first version. Then read the diff (cdk diff or terraform plan) before each apply, because an agent can open a security group to 0.0.0.0/0 just to make a connection test pass.

Add the checks to your workflow so they run on their own. The events platform's author used a Kiro hook to run Checkov before every push, then ran Checkov and Semgrep on each pull request. If you're on CDK, the AWS CDK MCP Server applies CDK Nag rules within the agent session.

Pro tip: Keep dev and production as two stacks deployed from the same code. If a fix only exists in production, the next agent-written change can overwrite it.

Step 6: Set up monitoring

Turn on Amazon CloudWatch logs and metrics before launch. ECS Express Mode sets up monitoring as part of the service, and Amplify Hosting reports hosting metrics with minimal delay, according to AWS.

Add alarms for error rates and latency, plus an external uptime check. Also set an AWS Budgets alert for Bedrock and the rest of your AWS usage.

Kiro credits need their own tracking in the Kiro console. Individual plans top up with prepaid add-on packs, and Enterprise overages stay off until an admin turns them on.

Step 7: Ship through a review gate

Protect the main branch so nothing deploys without an approval from the reviewer you named earlier. The pull request should carry the spec, the IaC diff, and the scanner results, so the approver judges the evidence along with the code.

This step only works once your organization decides who approves, and the governance section below covers that decision.

Security review before going live

Veracode's 2025 GenAI Code Security Report tested more than 100 LLMs in Java, Python, C#, and JavaScript. In those tests, 45% of code samples failed security tests and introduced OWASP Top 10 vulnerabilities. Run this review even if the scanners pass.

  • Exposed secrets: Scan the full Git history, including deleted files. GitGuardian detected 23.8 million new secrets in public GitHub commits in 2024. Fix by moving keys to Secrets Manager and rotating them. A pre-commit hook that blocks secret patterns can stop the next one.
  • Missing auth: Call every endpoint without a token. Any that answer with data are open to anyone who finds the URL. Fix with central middleware and role checks on the server, where a direct API call can't skip them.
  • Unsafe dependencies: Code models recommend packages that don't exist. A USENIX Security study of 576,000 code samples from 16 LLMs measured average package hallucination rates of at least 5.2% for commercial models and 21.7% for open-source ones, and attackers can register those names.

    To close that hole, verify each package exists, pin versions, and run software composition analysis.
  • No input validation: Look for SQL assembled with string concatenation and unescaped user input in the frontend. Also check for user text passed directly into a model prompt.

Each one has its own fix. Parameterized queries or an ORM handle the SQL, vetted sanitization libraries handle the frontend, and Bedrock Guardrails cover model-facing input.

In the same events platform, Checkov and Semgrep both passed a pull request, and the AWS Security Agent still flagged a high-severity issue. The change had removed anti-bot protection from a registration endpoint.

That's a logic flaw, and catching it takes a reviewer who knows what the endpoint should do.

Who approves AI-written code before release

In March 2026, the Financial Times reported that Amazon would require senior engineers to sign off on AI-assisted changes from junior and mid-level engineers, after a run of outages, according to The Decoder's summary from March 10, 2026.

Amazon disputed it on March 11, 2026. The company said the incidents were limited to its retail store infrastructure and that none involved AI-written code. It called reports of new approval requirements for engineers using AI tools "false."

The statement did name one AI-related incident, in which an engineer followed inaccurate advice that an AI tool inferred from an outdated internal wiki.

With Amazon's denial on record, your organization has to write its own approval rule, and it should exist before the first vibe-coded app goes live.

How to set the approval rule yourself

Write the rule down before the first app ships, and tie the level of review to what the change can break.

  • Changes that touch customer data, IAM, or payments need a senior engineer, while copy edits and internal dashboards can go through a standard review.
  • Approve the spec first and the diff second. With requirements.md and design.md settled, the code review checks the diff against a spec the reviewer already approved.
  • Keep the approver and the spec together with the scan results, so an incident six months later has a paper trail.
  • Give business teams guardrails by default. When the person building the app works outside engineering, those checks have to run on each change without a reviewer starting them.

Guardrails for business teams need tooling on top of the policy. Superblocks' production wall analysis takes the position that "Code generated by an agent is insecure until proven otherwise," and a business team without engineers will have a hard time providing that proof alone.

Superblocks 3.0, announced August 3, 2026, is aimed at business teams that create apps on AWS. Superblocks says it can deploy inside your own AWS VPC and route AI inference through Amazon Bedrock with the models your organization approves. Prompts and data stay in your AWS environment.

How much stays in your VPC depends on the deployment option. The full setup is the Enterprise Cloud-Prem option; Hybrid keeps production data in your VPC while Clark runs in Superblocks Cloud on non-production data.

Per the announcement, when an app needs storage, Superblocks provisions Aurora or S3 in that VPC and runs migrations when an app moves from dev to production.

The same announcement also describes the security layer. Superblocks says security agents and deterministic scanners check each code change for issues like SQL injection and hardcoded secrets, and that apps pull packages from a private registry with public npm blocked at the network level.

The App Database on AWS, in beta since June 1, 2026, runs on Amazon RDS with isolated dev and production environments. Rollbacks restore the app code and schema along with the data, and IT sets engine version, instance size, encryption, and backups centrally.

For the wider policy side, this vibe coding governance guide breaks the problem into 6 pillars, including a named owner for each app and data classification before any code is generated.

Common mistakes to avoid

Each mistake below undoes one of the steps above, so check this list again before the first production deploy.

  • Following a 2025 tutorial step by step. Q Developer closed to new subscriptions on May 15, 2026, and App Runner is closed to new customers. Check the AWS sunset list and the service's own docs before you depend on it, since App Runner's change sits on its availability page and not on that list.
  • Treating "it runs locally" as done. Local SQLite files, localhost URLs, and a .env that lives on one laptop can each break the first deploy to AWS.
  • Merging IAM policies the agent wrote without reading them. Wildcard actions and resources pass functional tests, so read each policy line by line.
  • Using Quick Spec on production work. It skips the approval gates that let a reviewer reject a plan at the design stage.
  • Ignoring usage-based costs. Kiro plan credits reset monthly, extra credits cost $0.04 each, and Bedrock bills per call. Budget alerts and credit limits belong in the initial deploy.

Vibe coding on AWS with Superblocks

Your team repeats the steps in Part 2 for each vibe-coded app. Superblocks says it applies the same controls at the platform level, so each new app starts with them in place.

That setup requires Enterprise. With Hybrid or Cloud-Prem deployment, a business team's app connects to your AWS account with secrets management, SSO, security scans, and audit logs set up once by IT.

What Superblocks handles for your team on AWS:

  • Keeps data in your AWS account (Enterprise): Cloud-Prem puts the full platform, including Clark’s AI inference, inside your AWS environment. Hybrid leaves your production data where it is and runs Clark on non-production data only.
  • Imports existing prototypes: Superblocks 3.0 imports apps from Claude, ChatGPT, Lovable, Replit, or raw code. On Enterprise, those imports then get checked by the platform's security agents and scans.
  • Gives IT a system of record: Audit logs and the Superblocks MCP, available to Enterprise customers, give admins programmatic access to builders, apps, integrations, permissions, and audit logs.
  • Works for non-engineers: Virgin Voyages reported 8 production apps in its first 30 days, made by 10 semi-technical builders after 2 days of training, and its customer story now counts 15+ apps.

One caveat applies to AWS teams. Virgin Voyages hosts Superblocks in Google Cloud, so treat it as evidence that business users can deliver production apps, and validate the AWS setup yourself.

Pricing: Teams starts at $125 per month billed monthly, or $100 per month billed annually. It includes 100 Governed Agent Units per month, up to 15 builders, and 1 hosted app, with each extra app at $10 per app per month. There's a 14-day free trial.

Enterprise uses custom pricing and adds VPC deployment (Hybrid or Cloud-Prem), SSO/SAML/OIDC, RBAC, audit logs, security agents, and a private package registry. Superblocks is SOC 2 Type II certified and HIPAA compliant.

To see an AI-generated app deploy inside your own AWS account, book a demo or start with the Superblocks introduction.

Frequently asked questions

Each answer below stands on its own and carries the source figure from this guide, so you can reuse it in an approval request.

Is AWS AI free?

No, AWS AI isn't free beyond starter limits. New AWS accounts get $100 in credits plus up to $100 more on a Free plan that lasts up to 6 months, and Bedrock usage draws from those credits. Kiro is billed separately, and its Free plan includes 50 credits per month.

What managed AI services does AWS offer?

AWS offers Amazon Bedrock for foundation models, Bedrock AgentCore for running agents, Amazon SageMaker AI to build, train, and deploy machine learning models, and Kiro for AI-assisted coding. Amazon Q Developer stays available in the AWS Management Console and in Slack and Microsoft Teams.

Is vibe coding a threat?

Vibe coding is a threat when AI-written code ships without review. In Veracode's 2025 tests, 45% of AI-generated code samples failed security tests and introduced OWASP Top 10 vulnerabilities.

Review is the control that addresses it. Scanners flag known patterns like hardcoded secrets and SQL injection, and a human reviewer is the check for logic flaws, such as a removed anti-bot control.

Does vibe coding count as coding?

Yes, vibe coding counts as coding because the output is source code that runs and breaks like any other, and it needs the same maintenance. An AI agent writes the code, and the person prompting owns the review and the deployment, along with every bug that reaches production.

Can Superblocks deploy vibe-coded apps on AWS?

Yes. Superblocks Cloud-Prem, an Enterprise option, runs the full platform inside your AWS environment, with AI inference through Amazon Bedrock. Teams starts at $125 per month billed monthly, but VPC deployment requires Enterprise's custom pricing.

Stay tuned for updates

Get the latest Superblocks news and internal tooling market insights.

You've successfully signed up

Request early access

Step 1 of 2

Request early access

Step 2 of 2

You’ve been added to the waitlist!

Book a demo to skip the waitlist

Thank you for your interest!

A member of our team will be in touch soon to schedule a demo.

8

production apps built

30

days to build them

10

semi-technical builders

0

traditional developers

8+

high-impact solutions shipped

2 days

training to get builders productive

0

SQL experience required

See full story →

See the full Virgin Voyages customer story, including the apps they built and how their teams use them.

Large cruise ship sailing in a harbor with a road lined with palm trees and cars in the foreground.
Why not Replit, Lovable, or Base44?

"Those tools are great for proof of concept. But they don't connect well to existing enterprise data sources, and they don't have the governance guardrails that IT requires for production use."

Superblocks Team
+2

Multiple authors

Nov 18, 2026