AWS Emulator puts a free copy of AWS on your computer, and AWS says AI agents wrote every line

A student learning to build apps on Amazon’s cloud usually needs an AWS account, a payment card and a steady connection. One wrong test can leave a bill. On 11 November, AWS released a free tool that removes all three, with an unusual claim about who built it.

The tool is AWS Emulator. An emulator is a program that copies how another system behaves, so your code can’t tell the difference. This one pretends to be AWS on your own computer. The bigger news: AWS says autonomous AI agents wrote all of it. Its team “hasn’t written or reviewed any of the code.” That raises a hard question for the software industry. Can software be written and shipped with no person reading it?

What the free AWS Emulator copies, and how developers use it

AWS says the emulator covers “more than 400 API models covering more than 200 AWS services”. That count comes from Swami Sivasubramanian, the AWS vice president for agentic AI; the official documentation gives no number. An API is the set of commands a program uses to talk to a service.

Developers point their usual AWS tools at it with a few settings; their code stays the same. No AWS account is needed, and AWS charges no fee; you pay only when you deploy on real AWS.

It is open source under CC0, a public domain dedication. It installs on Linux and runs on macOS and Windows through Docker. AWS does not accept code from the public; people report gaps on GitHub, where a list shows what does not work yet.

Excerpts from AWS’s announcement, as published on Amazon’s newsroom on 11 November. Image: Amazon Web Services

Three teams of AI agents, and a gate they must all pass

Matt Goldberg, a senior principal engineer at AWS, says it began as a weekend experiment: AWS’s Kiro agent copied Amazon SQS, a message service, in about 15 minutes. AWS now describes “thousands of autonomous agents” split into three teams. One writes tests and checks them against real AWS. A second runs them on the emulator and files failures. A third fixes them.

Every change must pass an automatic “merge gate”: it has to build, meet coding standards, survive security checks and keep matching real AWS. “Without this merge gate, we wouldn’t trust what our agents are building,” Goldberg writes. Sivasubramanian frames the human job as deciding “what gets built and when the work is correct”.

No one read the code, and AWS’s own warnings show the limits

AWS’s own documentation says not to use it for production, meaning live services customers use, or for security testing, because permission rules are not checked as real AWS checks them. Data is lost when it stops, and it is not tested for long or heavy workloads. Usage data, called telemetry, is sent to AWS by default, though you can switch it off.

How closely it matches each service varies; AWS publishes those scores with each release. On Hacker News, developers worried about long-term upkeep and about trusting code no human reviewed. A commenter saying they work at Amazon called it a real product; others wanted a formal public promise.

If you build on AWS, a free LocalStack alternative worth a careful look

The best-known local AWS emulator has been LocalStack. In 2026 it ended its free open-source Community Edition, InfoQ reported; it now needs an account, keeps a limited free plan for individuals, students and open-source work, and charges from US$39 per licence a month, billed yearly. Anyone wanting a free LocalStack alternative now has one from AWS itself. No independent test comparing AWS Emulator and LocalStack has been published yet.

Can you run AWS locally without an account? For testing, yes. For students and small teams, especially where dollar cloud bills hurt, that is a real saving. A sensible routine: test locally, check the “does not work” list, turn off telemetry if needed, and do a final check on real AWS.

Why trust matters more here than speed

On balance, this is a careful launch: free, open, honest about its limits, and measured against the real thing in public. In a test tool, a mistake usually shows up as a failed test, not leaked data. That makes it a sensible first home for unreviewed software.

The caution is real too: agents checking agents is a closed loop, and long-term upkeep is unclear. Our reading: this shows no-review software can work for a well-defined copying job, not that it is ready for systems handling money or health. For AWS, already spending heavily on AI, it is also a showcase.

AWS says it will share more at re:Invent. Watch two things there: a public promise to maintain the emulator, and whether its fidelity scores keep rising.