Kiro vs Claude Code: Specs First or Agent First?
November 6, 2026
Kiro vs Claude Code comes down to one choice: do you want the tool to make you write the plan down, or do you want an agent that just starts working and plans when you ask?
Kiro is AWS's agentic coding platform. It's built around specs: requirements, a design, and a task list, written before code changes. Claude Code is Anthropic's coding agent. It reads your codebase, edits files, and runs commands from your terminal, your IDE, a desktop app, or the browser. Kiro picks models from several labs. Claude Code runs Claude models only.
The short version:
- Pick Kiro if your problem is AI code that drifts from what you meant, you want requirements and designs saved in the repo, or your company buys through AWS
- Pick Claude Code if you live in the terminal, want one agent that works the same in the CLI, VS Code, JetBrains, desktop and web, and you already pay for a Claude plan
- Both have hooks, MCP, Agent Skills, subagents, and project instruction files, so extensibility is close to a tie
- Billing differs. Kiro meters credits per prompt. Claude Code draws from your Claude subscription's usage limits, or bills API tokens
Quick comparison
| Kiro | Claude Code | |
|---|---|---|
| Maker | AWS | Anthropic |
| Default workflow | Specs (requirements → design → tasks), or a "vibe" chat | Agent loop; plan mode on request |
| Main surfaces | Kiro IDE (Code OSS based), Kiro CLI, Kiro Web, mobile preview | Terminal CLI, VS Code and JetBrains extensions, desktop app, web, mobile |
| Models | Claude, OpenAI GPT, open-weight models, Auto router | Claude models (Anthropic, Bedrock, Google Cloud, Microsoft Foundry) |
| Project instructions | Steering files in .kiro/steering/, plus AGENTS.md | CLAUDE.md, or AGENTS.md, plus auto memory |
| Extensions | Hooks, MCP, skills, powers, custom agents, sub-agents | Hooks, MCP, skills, plugins, subagents |
| How you pay | Free tier, then credit plans; credits weighted by model | Included in paid Claude plans, or pay per API token |
Spec-driven vs agent-first
Take a task like "add rate limiting to the public API."
In Kiro, you'd start a spec. Kiro's spec workflow writes a requirements file with user stories and acceptance criteria, then a design file, then a task list. You approve each one before the next is built. After that, Kiro works through the tasks and shows each as in progress or done. The three files stay in the project, so the reasoning behind the code is still there next month.
In Claude Code, you'd just describe the change. It reads the relevant files, writes code across them, runs your tests, and shows you the diffs. For a change you want to inspect first, you press Shift+Tab into plan mode. Anthropic's docs describe it plainly: Claude "reads files and proposes a plan but makes no edits until you approve."
So both can plan. The difference is what the plan becomes:
| Kiro spec | Claude Code plan mode | |
|---|---|---|
| When it happens | The main path for feature work | When you switch into it |
| What you get | Three files: requirements, design, tasks | One proposed plan you approve or edit |
| Lives where | In the repo, next to the code | In the session, unless you ask Claude to save it |
| Task tracking | Live status per task | The agent follows the plan and reports back |
Neither tool forces one style. Kiro has a plain chat for quick work, and Claude Code can write a spec document if you ask it to. The defaults are what you'll live with, though. Kiro asks "what are we building?" first. Claude Code asks "what should I do?" and gets going.
Interface: Kiro IDE and CLI vs Claude Code everywhere
This is the biggest practical difference.
Kiro started as an IDE. It's based on Code OSS, so you can import VS Code settings, themes, and Open VSX plugins. Specs, steering, and hooks show up in panels in the editor. Then there's Kiro CLI, which has a history worth knowing. Kiro's Q CLI upgrade guide says "Kiro CLI is the next update of the Q CLI." Amazon Q Developer CLI users were moved to it in November 2025. The license changed too: Q Developer CLI was Apache 2.0, and Kiro CLI is under the AWS Intellectual Property License. In other words, the CLI is no longer open source. Kiro CLI 3.0 runs the same agent harness as the IDE and Kiro Web, so your steering, hooks, and MCP servers carry across. There's also Kiro Web for cloud sessions that keep running after you close your laptop.
Claude Code started in the terminal and spread out from there. The same agent now runs in the CLI, extensions for VS Code (which also install into Cursor) and JetBrains IDEs, the Claude desktop app, and the browser at claude.ai/code. The CLI is built for scripting: you can pipe logs into it, run it headless with claude -p in CI, or wire it into GitHub Actions.
Who each one suits:
- You want a full editor with the agent built in → Kiro IDE
- You want to keep your current editor → Claude Code's extension, or either CLI
- You script agents in CI → both work; Kiro CLI has headless runs, and Claude Code adds
claude -pand GitHub Actions - You were on Amazon Q Developer CLI → you're already on Kiro CLI. If you're reviewing the rest of Q Developer, see our Amazon Q Developer alternatives guide
Models
Kiro is not tied to one lab. Its models page lists Claude models, OpenAI's GPT models, and open-weight models such as DeepSeek, MiniMax, GLM, and Qwen. By default, an Auto router picks a model per task. Each model has a credit multiplier, so a premium model costs more credits than Auto for the same prompt.
Claude Code uses Claude models only. You can switch between Opus, Sonnet, and Haiku with /model, and the opusplan setting uses Opus for planning and Sonnet for the build. Companies can run it through Amazon Bedrock, Google Cloud, or Microsoft Foundry instead of Anthropic directly.
If you specifically want Claude, both tools give it to you. Kiro just gives you other options too. Model lineups change monthly, so check each vendor's current list.
Pricing model
The two bill in very different ways.
Kiro uses credits. Its pricing page says "A credit is a unit of work in response to user prompts." There's a free tier with a small monthly credit allowance and limited models. Paid individual plans add more credits and premium models, and you can buy add-on credits when you run out. Team plans pool credits across seats. Spec task runs and hook runs also use credits.
Claude Code comes with a Claude plan. Per Claude's pricing page, Claude Code is included in all paid plans. It isn't in the free plan. Your terminal work and your chats draw from the same usage limits, which reset on a rolling five-hour window plus weekly caps. You can also skip the subscription and pay per API token through an Anthropic Console account.
Which is cheaper depends on how you work:
- Light or bursty use → Kiro's free tier or a low credit plan, since you only spend on prompts you make
- Heavy daily use, mostly Claude → a Claude subscription is often the better deal, and you get Claude chat as well
- A team on AWS billing → Kiro's pooled credits fit existing AWS procurement
Our Claude Code usage limits guide covers how the subscription caps work in practice.
Extensibility: hooks, steering, MCP, skills, subagents
The two tools have converged here more than anywhere else.
| Feature | Kiro | Claude Code |
|---|---|---|
| Project instructions | Steering files (product.md, tech.md, structure.md by default) with inclusion modes | CLAUDE.md, read at the start of each session; auto memory saves learnings |
| AGENTS.md | Supported | Read in place of CLAUDE.md |
| Hooks | JSON files in .kiro/hooks/; run a shell command or inject a prompt | Settings-based hooks on dozens of events; run a command, HTTP call, MCP tool, prompt, or agent |
| MCP | Across IDE, CLI, and web | Local and remote servers |
| Skills | Agent Skills standard, in .kiro/skills/ | Agent Skills standard, invoked automatically or as /skill-name |
| Subagents | Custom agents can run as sub-agents in parallel | Subagents in .claude/agents/ with their own context and tools |
| Packaging | Powers bundle tools, MCP, and know-how per technology | Plugins bundle skills, agents, hooks, and MCP servers |
A few differences matter:
- Kiro's steering ships with a template. Three default files tell the agent what the product is, which tech it uses, and how files are laid out. Claude Code starts empty, though
/initdrafts a CLAUDE.md from your repo. - Claude Code's hooks cover more events, including permission requests, subagent start and stop, and context compaction. Kiro's hooks focus on file saves, tool calls, and task completion, and you can create them by asking the agent in chat.
- Skills are portable. Both follow the open Agent Skills standard, so a skill you write for one usually works in the other.
When to pick each
| Your situation | Pick |
|---|---|
| Features with fuzzy requirements | Kiro |
| Requirements and designs need to live in the repo for review | Kiro |
| Company standardizes on AWS identity and billing | Kiro |
| You want models from several labs | Kiro |
| You work in the terminal and script agents | Claude Code |
| You want the same agent in your current editor, desktop, and browser | Claude Code |
| You already pay for Claude Pro or Max | Claude Code |
| Many small fixes all day | Claude Code |
For how both stack up against an AI editor, see Kiro vs Cursor, which goes deeper on Kiro's steering and team features, and Claude Code vs Cursor.
Using both
You don't have to choose. Because both read AGENTS.md and both use the Agent Skills format, one repo can serve both tools:
- Put shared rules in AGENTS.md. Both agents load it.
- Plan in Kiro. Let it write requirements, design, and tasks for a big feature.
- Build with Claude Code. Point it at the tasks file and have it work through the list from your terminal or editor.
- Keep skills in one format. Copy them between
.kiro/skills/and.claude/skills/.
The cost is two bills and two sets of settings, so try it on one feature before you commit to both.
When the setup is the problem
Kiro and Claude Code both assume you have a repo and know your way around a terminal or IDE. A lot of people get stuck before that: they found a working AI app or workflow and can't get it running. Taku is an AI-native desktop workspace where you mirror a setup someone already got working, run it, and remix it, without building their environment yourself. Taku is in Beta, and the Mac app is available now.
FAQ
Is Kiro better than Claude Code?
Neither is better across the board. Kiro is better when you want requirements and a design written down before code, and when you want models from more than one lab. Claude Code is better for fast, terminal-first work and for using one agent across your CLI, IDE, desktop, and browser.
Does Kiro use Claude?
Yes. Kiro offers Claude models alongside OpenAI GPT models and open-weight models, with an Auto mode that picks one per task. Claude Code uses only Claude models.
Is Kiro CLI the same as Amazon Q Developer CLI?
Kiro CLI is the successor. Kiro's docs call it "the next update of the Q CLI," and Q Developer CLI users were upgraded to it starting in November 2025. Existing workflows, subscriptions, and .amazonq project settings keep working, but Kiro CLI is under the AWS Intellectual Property License rather than Apache 2.0.
Does Claude Code have spec-driven development?
Not as a built-in mode. Plan mode makes Claude read your code and propose a plan you approve before any edits. You can ask Claude Code to write requirement and design documents, but it won't generate and track three linked spec files the way Kiro does.
How does Kiro vs Claude Code pricing compare?
Kiro charges credits per prompt, with a free tier and paid credit plans, and premium models use more credits. Claude Code is included in paid Claude plans and shares their usage limits, or you pay per API token. Check Kiro's pricing and Claude's pricing page for current plans.
Can I use Kiro and Claude Code in the same project?
Yes. Both read AGENTS.md and support Agent Skills and MCP, so you can share rules and skills. A common split is planning in Kiro and building with Claude Code.