Kiro is an agentic IDE from Amazon Web Services built on a VS Code core and powered by Claude Sonnet 4 via Amazon Bedrock. Its defining characteristic is a spec-driven development workflow: before writing a single line of code, Kiro converts a natural-language feature request into a formal requirements specification, a design document, and a structured task list, giving teams an auditable engineering artifact alongside the implementation. More than 250,000 developers joined the public preview within weeks of launch, signaling broad industry interest in structured agentic workflows.
Key capabilities
Spec-driven development pipeline — Kiro’s core loop starts with the developer describing a feature in plain language. The agent produces a requirements spec and a design document, requests approval, then decomposes the work into discrete tasks before touching any source file. This means every Kiro session produces traceable documentation as a byproduct of normal coding.
Amazon Bedrock and Claude 4 integration — Kiro routes all inference through Amazon Bedrock, giving enterprise users the access controls, audit logging, and data residency guarantees that Bedrock provides. The model backing agent tasks is Claude Sonnet 4, one of the strongest general-purpose coding models available as of mid-2026.
Kiro Powers domain extensions — AWS has partnered with Stripe, Figma, and Datadog to ship domain-specific extensions called Kiro Powers, which give the agent pre-built knowledge of each platform’s APIs, webhooks, and common integration patterns. This reduces hallucination risk when working with external services.
Web service availability — In addition to the downloadable desktop IDE, Kiro is available as a web application at app.kiro.dev, lowering the barrier for teams that cannot install local tooling or want to use Kiro from a browser or a managed development environment.
Autonomy level
Kiro operates at autonomy level 4. After receiving a natural-language description, it autonomously generates the full spec, design, and task breakdown, then executes implementation tasks sequentially, running tests and iterating on failures without requiring per-step prompting. Human approval gates are placed at the spec review stage and at the completion of significant task groups, keeping the developer in control of the overall direction while the agent handles execution detail.
Strengths
- Spec-first workflow produces engineering documentation as a natural byproduct of coding
- Deep AWS and Amazon Bedrock integration satisfies enterprise compliance and data residency requirements
- Kiro Powers extensions bring domain-specific API knowledge for Stripe, Figma, and Datadog out of the box
- Claude Sonnet 4 provides top-tier reasoning on complex multi-file tasks
- Free tier of 50 interactions per month is accessible for individual developers evaluating the approach
- Available as both a downloadable IDE and a web service at app.kiro.dev
Limitations
- Closed source with no community contribution path to the core agent or spec engine
- Full Bedrock-backed features require an AWS account, which adds onboarding friction for non-AWS shops
- Spec-driven workflow is also a constraint: developers who prefer rapid, exploratory iteration may find the mandatory planning phase slows them down
- Public preview status means the tool is relatively new and feature completeness is still evolving
- Model support is limited to Claude via Bedrock; no option to bring your own inference endpoint
- No offline or air-gapped operation mode