Note: This overview synthesizes official Kiro documentation and third-party coverage. I have not personally tested Kiro CLI for an extended period. Treat feature descriptions as what the product claims to offer, not verified performance.
The AI coding CLI space keeps expanding. Claude Code, Cursor's terminal mode, Aider, Codex CLI. Most share a common limitation: every session starts fresh. You explain project conventions, testing preferences, naming standards. Then repeat tomorrow.
Kiro, built by a team within AWS, takes a different approach. The CLI version brings the IDE's steering files and spec-driven development into the terminal. It went into general availability on November 17, 2025, alongside team features and AWS IAM Identity Center support.
What Kiro CLI Actually Is
Kiro CLI evolved from Amazon Q Developer CLI. According to the migration documentation, existing Q Developer CLI users with auto-update enabled were automatically upgraded to Kiro CLI on November 24, 2025. The underlying agentic capabilities carry over: agent mode, MCP support, steering, and custom agents.
The installation is straightforward:
curl -fsSL https://cli.kiro.dev/install | bash
Kiro CLI runs on macOS and Linux (including AppImage and Ubuntu). No Windows support currently.
Steering Files
The core concept that differentiates Kiro from other CLI tools: markdown files in .kiro/steering/ (project-level) or ~/.kiro/steering/ (global) that provide persistent instructions.
Unlike Claude Code's single claude.md or Cursor's unified rules file, Kiro's approach is modular. The documentation describes three foundational steering files that Kiro auto-generates:
- structure.md: codebase architecture
- tech.md: technical stack and patterns
- product.md: business context and requirements
You can add custom steering files for domain-specific knowledge: logging standards, security policies, testing conventions. According to AWS re:Invent 2025 presentations, best practices include organizing documents by domain, using 3-5 documents per task, and treating steering files as version-controlled code.
The practical appeal: if you maintain multiple codebases with shared conventions, writing instructions once and reusing them beats rebuilding context for each project. Teams can collaborate on steering files and share them across projects.
Spec-Driven Development
Kiro's other major differentiator is specifications. Specs are structured markdown files that formalize feature development through three phases:
requirements.md: User stories with acceptance criteria in EARS notation (Easy Approach to Requirements Syntax, originally developed at Rolls Royce). Format:
WHEN [condition/event] THE SYSTEM SHALL [expected behavior]design.md: Technical architecture, sequence diagrams, implementation considerations
tasks.md: Discrete, trackable implementation steps
The workflow: describe a feature, Kiro generates requirements, you refine them, Kiro produces a design document, then breaks it into tasks. Each task can be executed individually with progress tracking.
This approach explicitly separates planning from coding. According to AWS, addressing issues during planning is 5-7x less costly than resolving them during implementation. Whether that holds true for AI-assisted development specifically is an open question.
Models and "Auto" Agent
Kiro uses Claude models under the hood. The pricing page lists available options:
- Auto (default): A mix of frontier models (Sonnet 4.5 and specialized models) with intent detection and caching to balance quality, latency, and cost
- Sonnet 4: Direct access, costs 1.3x more credits than Auto
- Sonnet 4.5: Available
- Haiku 4.5: Available
- Opus 4.5: Available
The Auto agent is the key differentiator from just using Claude directly. It routes prompts to appropriate models and applies Kiro-specific optimizations.
Pricing: The Complicated Part
Kiro's pricing has been controversial. The Register reported user complaints about the credit system being opaque and consuming more requests than expected.
Current tiers from the pricing page:
| Tier | Monthly Cost | Credits | Overage |
|---|---|---|---|
| Free | $0 | 50 | None |
| Pro | $20 | 1,000 | $0.04/credit |
| Pro+ | $40 | 2,000 | $0.04/credit |
| Power | $200 | 10,000 | $0.04/credit |
New users get 500 bonus credits usable within 30 days.
Credits are the tricky part. A credit is "a unit of work in response to user prompts." Simple prompts can consume less than 1 credit. Complex prompts (like executing a spec task) typically cost more than 1 credit. Using Sonnet 4 directly costs 1.3x more credits than Auto for the same task. Credits meter to the second decimal point (minimum 0.01 credits).
What consumes credits: vibe mode prompts, spec mode prompts, spec refinement, task execution, and agent hook execution.
Early user reports suggest the credit consumption can be unpredictable. One GitHub issue claimed Pro+ limits "were completely consumed within 15 minutes of usage in a single chat session." Your experience will depend heavily on prompt complexity and which modes you use.
How It Compares
| Aspect | Kiro CLI | Claude Code | Aider |
|---|---|---|---|
| Instruction approach | Multiple steering files | Single claude.md | Per-session context |
| Spec generation | Three-phase structured workflow | Not specialized | Not specialized |
| AGENTS.md support | Yes | Yes | Partial |
| Model selection | Auto + Claude options | Claude models | Multiple providers |
| Pricing model | Credit-based tiers | $20/mo subscription | Free + API costs |
| Platform | macOS, Linux | macOS, Linux, Windows | Cross-platform |
The spec-driven workflow is genuinely different from competitors. Whether that structured approach fits your workflow depends on project complexity and team size.
What's Missing From This Overview
I haven't tested:
- Actual credit consumption patterns across different project types
- Real-world steering file behavior and precedence rules
- Spec generation quality on legacy codebases
- MCP integration reliability
- Performance compared to Claude Code on equivalent tasks
- Team collaboration features
The pricing controversy suggests the tool works well enough that people want to use it heavily, but the credit system creates friction. That's a very different problem than "the tool doesn't work."
Should You Try It?
If you're frustrated by repeating project context across AI coding sessions, Kiro's steering file approach addresses a real problem. The modular organization beats a single instruction file, especially for teams with shared conventions.
The spec-driven development workflow makes sense for complex features where planning matters. For quick fixes or exploration, the overhead might not be worth it.
Before committing:
- Start with the free tier (50 credits + 500 bonus)
- Test steering file behavior on a small project
- Track your credit consumption carefully before upgrading
- Compare against Claude Code's $20/month flat rate for your usage pattern
The product is real, actively developed, and backed by AWS. Whether the credit-based pricing model works for your usage pattern is the main question.
