Overview
AWS Emulator is a tool released by AWS on November 11, 2026 that gives coding agents a local copy of AWS to test against, released under CC0 and hosted on GitHub.
The problem it solves is concrete. An agent works in a loop: write code, run it, read the result, decide what to change. In distributed systems that loop has to run inside infrastructure. Real AWS means an account, quotas, a bill and network round trips, so a single cycle costs seconds to minutes. A local copy collapses that loop onto the machine, letting the agent check its own work while writing.
In scope it is built from more than 400 Smithy API models covering over 200 AWS services. Smithy is the public format AWS uses to describe a service's operations, inputs, outputs and errors, so the boundaries come from published interface definitions rather than guesses.
How it was built is more interesting than the product. The idea came from Matt Goldberg, a Senior Principal Engineer on the AWS Developer Tools team: given a written description of how a service should behave and the real service to check against, an agent should be able to build a working implementation. The method starts from fidelity tests, end-to-end customer scenarios derived from the public API models, for example verifying that a received SQS message stays hidden from other readers until its visibility timeout expires. Agents first get each test passing against real AWS, which turns the test into a record of real behaviour, then build the emulator until it passes the same test. A test that still fails points at a gap.
It started small. Matt tried SQS alone, gave Kiro one prompt describing the approach, and got a passable first version in about 15 minutes. An internal version came together over a weekend, gained adoption within a month, and turning it into a robust solution took far more work.
The merge gate deserves its own paragraph. Every change passes automated checks that include multiple security agents, each reviewing for a different way to attack the change, alongside a general code reviewer and deterministic checks that stop agents touching code they should not. The gate is itself a sample of agent engineering: when agents write the code, the engineer's job moves to specs, tests and gates.
The comparison case makes the value of specs and a test environment vivid. Werner Vogels rebuilt a replication engine that took about four months the first time. He supervised the agent for four nights before realising it had no way to check its results; once it had a spec and a test environment, the rebuild took about two hours.
Known gaps are not hidden. A things-that-do-not-work folder in the repository holds the issues agents found where behaviour does not yet match real AWS, and both commit history and issues are public. Reading that folder first beats reading any announcement.
Be clear about what this is. It is a test environment, not a production replacement and not a complete service implementation. Fidelity reaches exactly as far as the fidelity tests do, and anything uncovered is blank. Account-free, quota-free and offline, it fits CI and teaching. AWS has said it will share more at re:Invent, which signals the thing is still moving fast.
Key Features
- Runs locally and offline: Executes on your machine without connecting to real AWS, needing no account or network, which turns the agent's self-check from a network round trip into local execution.
- Over 200 services: Built from more than 400 Smithy API models, with boundaries taken from the interface definitions AWS publishes covering operations, inputs, outputs and errors.
- Fidelity-test driven: Tests are end-to-end scenarios derived from public API models, first made to pass against real AWS to record real behaviour, then required to pass against the emulator, so failing tests point at gaps.
- Released under CC0: Dedicated to the public domain, usable, modifiable and commercial without attribution.
- Written entirely by agents: All code was written and tested by agents, from a 15-minute first SQS version to over 200 services, with commit history and issues public.
- Merge gate with security agents: Every change passes automated checks where several security agents each probe a different attack surface, alongside a general reviewer and deterministic checks limiting what agents may touch.
- Public gap list: A things-that-do-not-work folder records behaviours known to differ from real AWS, so you can judge coverage before relying on it.
- No account, no bill: No call charges or quotas, which makes it usable in CI, offline environments and teaching.
Use Cases
- Local self-checking and regression for coding agents
- Development and CI where no cloud account exists
- Verifying multi-service orchestration logic
- Learning and teaching distributed system behaviour
- Offline demos and sandboxes
- A merge gate for agent-written code
Pros
- Free and offline on your machine, so agents self-check without connecting to AWS and iterations do not wait on environments or bills
- More than 400 API models covering over 200 services, two orders of magnitude beyond hand-built mocks
- Fidelity tests calibrated against real AWS, making each test a record of behaviour and each failure a located gap
- CC0 with public commit history and issues, plus a readable gap list
- Entirely agent-written and gated by multiple security agents, a legible sample of agent engineering
- No call charges or quotas, so it fits CI and offline environments
Pricing
Free. Code is released under CC0, allowing commercial use and modification without attribution. Hosted on GitHub with public commit history and issues. AWS has said it will share more at re:Invent.
Summary
The hardest part of a coding agent is not writing code, it is having no way to verify what it wrote because the cloud services it calls do not exist on the machine. AWS Emulator moves that layer onto the laptop: free, offline, covering more than 200 services. Its value is not simulation fidelity but letting an agent close the write, run, read loop without waiting on a real environment or a bill. The more interesting part is how it was built: agents wrote all of it and calibrated its tests against real AWS, with known gaps published in the repo.
Version History
- AWS Emulator released (2026-11-11): AWS released AWS Emulator, giving coding agents a local offline copy of AWS built from more than 400 Smithy API models covering over 200 services, released under CC0. The whole implementation was written and tested by agents: fidelity tests are derived from public API models, made to pass against real AWS first, then required to pass against the emulator. Matt Goldberg produced a first SQS emulator in about 15 minutes from a single Kiro prompt. Every change passes a merge gate with multiple security agents doing adversarial review plus deterministic checks. Known divergences from real AWS are listed in the repository's things-that-do-not-work folder.