The tool relies on a structured approach called “spec-driven development”. Before generating a single line of code, it establishes detailed specifications, a complete technical architecture and a sequenced action plan.
It’s a little ghost who is starting to make people talk about him. Developed by AWS teams and used internally at Amazon, Kiro, this new IDE powered by AI, relies on spec-driven development to generate more precise code, supposed to directly respond to the expectations of developers and software engineers. The tool is at the heart of the AWS strategy to retain developers, almost a loss leader to attract them to the AWS ecosystem. Largely pushed internally, it benefits from particular attention which suggests solid support over time. But what is it really worth? Should we adopt it? Analysis and testing.
An IDE focused on spec-driven development
Kiro’s particularity can be summed up in three words: spec-driven development. Where most AI code agents are content to generate code on the fly without a plan (less and less certainly), Kiro imposes a more structured approach. Concretely, when a user submits a request in natural language, the tool does not immediately generate the code. It begins by converting the query into a sort of detailed specification, then produces a complete technical architecture (flow diagrams, database diagram interfaces, etc.), before breaking everything down into sequenced tasks. It is only after validation of the specifications (by the user) that the code is generated. In short, a sort of Claude Code planning mode but improved. The objective is to avoid endless back and forth and produce code that really meets initial expectations.
On the model side, Kiro is banking exclusively on the Claude family from Anthropic, a choice which is not insignificant when we know that Amazon has injected several billion dollars into the start-up. Developers have access to Claude Sonnet 4.0 and 4.5 for common coding tasks, Claude Opus 4.5 for complex projects requiring advanced reasoning, and Claude Haiku 4.5 for when speed trumps power.
Plans from 0 to 200 dollars per month
In terms of prices, Amazon plays the flexibility card with a credit system. Each interaction with the AI, whether it is a simple prompt, a spec generation or the execution of a task, consumes credits, the cost of which varies depending on the complexity of the request and the model requested. A prompt processed by Auto mode (which intelligently mixes the models) will cost less than the same prompt sent directly to Sonnet 4 for example. In detail, a prompt consumes approximately 1 credit in auto, 1.3 credits with Sonnet 4.5 or Sonnet 4, 0.4 credits with Haiku 4.5 and 2.2 credits with Opus 4.5 (the largest model).
The free plan offers 50 monthly credits. For developers, three paid subscriptions are offered: Pro at $20 per month for 1,000 credits, Pro+ at $40 per month for 2,000 credits, and Power at $200 per month for 10,000 credits. Finally, additional credits are billed at $0.04 each. The tool is available on macOS, Windows and Linux.
|
Formula |
Monthly price |
Credits included |
Additional credits |
|---|---|---|---|
|
Free |
Free |
50 |
Not available |
|
Pro |
$20 |
1,000 |
$0.04/credit |
|
Pro+ |
$40 |
2,000 |
$0.04/credit |
|
Power |
$200 |
10,000 |
$0.04/credit |
The JDN test
Kiro sets up quite quickly and feels quite light in use. Likewise, the general interface is quite simple and rather intuitive, unlike Google’s competing solution (Antigravity), although also based on a fork of VS Code. The interface is clear, with a code editor and the Kiro agent within the same, practical interface (unlike, again, Antigravity).
For this test, we will ask Kiro to generate a chess game as a web widget with a computer based on the Minimax algorithm (a decision method used in two-player games). With Kiro, it is possible to use two modes for the AI agent: Vibe which allows you to create code as soon as the prompt is sent and Spec which allows you to create a plan and technical specifications before the AI generates the code. We will use the latter mode.
Prompt:
Crée un jeu d'échecs en web app sous forme de widget web, tout côté client sans backend. Le joueur humain joue les blancs, l'ordinateur joue les noirs. Le plateau 8x8 doit être interactif avec des pièces cliquables. Toutes les règles officielles doivent être respectées : déplacements légaux de chaque pièce, prises, roque petit et grand, prise en passant, promotion du pion en dame. Le jeu doit détecter les situations d'échec, d'échec et mat, et de pat. L'IA adverse utilise l'algorithme Minimax avec une fonction d'évaluation basée sur la valeur matérielle des pièces. Affiche l'historique des coups joués en notation algébrique et ajoute un bouton pour recommencer une nouvelle partie.
Kiro then begins creating the specs and proposes a document containing an implementation plan. The user can then modify it to suit their needs and specify specific libraries to use or additional functions.
Once the specs have been edited, it is time to start generating the code. First remark: the scaffold around the model developed by the Amazon teams makes the agent particularly intelligent. The agent goes on the web to look for best practices for implementing a Minimax engine in a widget and adapts automatically. It then generates a design document (design.md) which summarizes the main implementations that the agent wishes to put in place with all the elements that will be defined. The file describes many of the classes that will be used.
Once the design document is accepted, Kiro then generates the list of tasks to be carried out to generate the project as a whole. It offers two modes: a quick MVP mode and a full approach mode. The first allows you to eliminate non-essential tasks to quickly get to the final product and the last allows you to generate the entire project at once (documentation, unit test, etc.). We will choose the complete option with all the tasks.
We finally launch the code generation. Each of the 11 tasks must be launched by hand, which allows fine control but still requires constant human attention. The process is quite long and time-consuming because the AI systematically runs tests to verify that the new features implemented have not broken anything. On a large project, this can take time and consume a lot of credits, but guarantees near-production code quality. On the other hand, for simple experimentation, it is probably too much. For our chess game, Kiro will have used approximately 83 credits. It’s a lot, but our game is fully functional. Everything works perfectly: the opposing AI, the commands, the move history… Kiro’s spec-driven development has a real advantage in producing almost production-ready code.
A good alternative to Cursor and Windsurf
Kiro clearly stands out from other AI-powered IDEs with its approach truly designed for demanding developers. The controls are finer, the process more rigorous. We are much closer here to software engineering than pure vibe coding, and that’s so much the better. For teams that need precision on large and complex projects, this is the ideal tool. A real alternative to Cursor and Windsurf which deserves our attention.
Kiro also has a CLI agent (a dedicated article would be needed to test it in depth) which completes the offer and makes it a real ecosystem. Does he compete head-on with Claude Code? Not really, because it is less universal in its approach, but it is clearly designed for software engineers who want to keep control of their code. Kiro is, in our opinion, an IDE that deserves to be adopted even if the credits go very quickly. Amazon has obviously invested in its product, and the result is there.
