What is the difference between Codex and Kiro?
Codex is OpenAI's coding agent that runs in your terminal, editor, desktop app and the cloud, while Kiro is an agentic IDE from AWS that writes specs (requirements, design and tasks) before it codes. Codex is conversation first with a built-in sandbox; Kiro is spec first, built on Code OSS, and lets you pick from several model vendors.
Both read AGENTS.md, support Agent Skills, MCP servers and hooks, and both have a terminal agent and a way to run tasks remotely. The differences are in workflow, plans and safety defaults. Everything below comes from official docs read on 26 October 2026.
How does Codex work?
The Codex repository describes the CLI as a "Lightweight coding agent that runs in your terminal," licensed under Apache 2.0. You install it and run it in a project:
# macOS and Linux
curl -fsSL https://chatgpt.com/codex/install.sh | sh
# or with npm or Homebrew
npm install -g @openai/codex
brew install --cask codex
cd your-project
codex
On first run you choose Sign in with ChatGPT or use an API key. Beyond the CLI, Codex runs as an IDE extension, inside the ChatGPT desktop app and as Codex Cloud, which works on remote machines from published environments. The Codex CLI docs list codex exec for scripts, codex resume for old chats and codex cloud for sending work to the cloud.
For planning, the CLI commands reference lists /plan, which switches the chat to plan mode so Codex proposes "an execution plan before implementation work starts." The plan lives in the conversation rather than in a fixed set of files.
Codex runs on OpenAI models by default. The models page recommends gpt-6.1-sol on paid plans when available, with near Astra performance at a lower cost than gpt-6-astra, and gpt-6-luna for focused, repeatable tasks. The CLI can also use custom model providers, or local models through Ollama or LM Studio with --oss, per the advanced config docs.
How does Kiro work?
The Kiro site says it is "Built and operated by AWS" with the tagline "Move beyond AI coding to agentic engineering." The FAQ adds: "Kiro works as an agent in the IDE, with a CLI, in a web browser, from mobile app, and in Kiro Crew." The IDE "is based on Code OSS," so VS Code settings, themes and Open VSX compatible extensions carry over.
Kiro's signature feature is specs: requirements.md (or bugfix.md for bugs), design.md and tasks.md, stored under .kiro/specs/ and meant to be committed. Requirements use EARS notation (WHEN a condition, THE SYSTEM SHALL do something), and the agent then works through the task list. See the Kiro IDE guide for the full walkthrough.
The terminal agent installs with one script, per the Kiro CLI docs:
curl -fsSL https://cli.kiro.dev/install | bash
cd my-project
kiro-cli
On models, the FAQ says "By default, your prompts are processed by Auto, an agent that uses a mix of different frontier models." You can also choose Anthropic's Claude models, "OpenAI's GPT-5.6 (Sol, Terra, and Luna)" or open weight models such as DeepSeek 3.2, MiniMax M2.5, GLM-5 and Qwen3 Coder Next. You do not need an AWS account; sign in with GitHub, Google, AWS Builder ID or IAM Identity Center.
How do Codex and Kiro compare side by side?
| Codex | Kiro | |
|---|---|---|
| Maker | OpenAI | AWS |
| Main surfaces | CLI, IDE extension, ChatGPT desktop app, Codex Cloud | Code OSS based IDE, CLI, web, mobile, Kiro Crew |
| Planning style | Conversation, with /plan mode |
Specs: requirements.md, design.md, tasks.md |
| Models | OpenAI by default, such as gpt-6.1-sol and gpt-6-luna; custom or local providers in the CLI |
Auto, Claude, GPT-5.6, open weight models |
| Lowest paid plan (Oct 2026) | ChatGPT Go, $8 a month (desktop app only); Plus, $20 a month, adds CLI, IDE and cloud | Kiro Pro, $20 per user a month |
| Default safety | OS sandbox, workspace-write with on-request approvals in Git folders, network off |
Reads allowed in the workspace, most other actions prompt |
| Project instructions | AGENTS.md |
Steering files in .kiro/steering/, plus AGENTS.md |
| Agent Skills | .agents/skills/, ~/.agents/skills/ |
.kiro/skills/, ~/.kiro/skills/ (global skills in IDE and CLI only) |
| MCP config | [mcp_servers] in ~/.codex/config.toml |
.kiro/settings/mcp.json, ~/.kiro/settings/mcp.json |
| Hooks | hooks.json or [hooks] in config.toml |
JSON files in .kiro/hooks/ |
| Scripting | codex exec |
kiro-cli chat --no-interactive with KIRO_API_KEY |
| Remote tasks | Codex Cloud, GitHub repositories | Kiro Web, GitHub or GitLab, us-east-1 only |
How do they handle permissions and safety?
Codex puts a sandbox around the agent. The approvals and security docs say "By default, the agent runs with network access turned off," and locally Codex "uses an OS-enforced sandbox that limits what it can touch." In a version-controlled folder it recommends the Auto preset (workspace-write plus on-request approvals), so it edits and runs commands in the project and asks before going further.
Kiro uses permission rules. The permissions docs say the default policy allows reading files in the workspace and some read-only git and system commands, and "Everything else prompts for approval." You add allow, ask and deny rules in permissions.yaml, and a deny rule always wins. In the IDE, Agent Autonomy switches between Autopilot, where the agent "proceeds with allowed operations without prompting," and Supervised, where it "prompts before any action." Kiro Web runs in a cloud sandbox instead of prompting per action.
Both approaches keep you in the loop by default. Codex leans on OS-level isolation; Kiro leans on explicit rules you can tune per command.
How do instructions, skills, MCP and hooks compare?
Instructions. Codex reads AGENTS.md from your Codex home and from the project root down to your working folder, capped at 32 KiB by default. Kiro's steering docs use Markdown files in .kiro/steering/ with inclusion modes (always, fileMatch, manual, auto), and say "AGENTS.md files do not support inclusion modes and are always included." One AGENTS.md at the repo root therefore steers both tools. The AGENTS.md guide covers the format.
Skills. Both follow the Agent Skills standard. Kiro's skills docs require a name that matches the folder (lowercase letters, numbers and hyphens, max 64 characters) and a description of up to 1,024 characters, the limits of the Agent Skills standard. The folders differ: .kiro/skills/ for Kiro and .agents/skills/ for Codex. See Codex skills.
MCP. Codex stores servers as [mcp_servers.<name>] tables in config.toml. Kiro uses an mcpServers object in mcp.json, merging user and workspace files with workspace taking precedence. The formats differ, so configure each tool separately.
Hooks. Kiro's hooks docs say "Hooks run shell commands or agent prompts automatically when specific events happen in your session," with triggers such as PostFileSave, PreToolUse, PostToolUse and Stop. Codex hooks cover similar events, including PreToolUse, PermissionRequest and Stop, and need a trust review before they run. The Codex hooks guide has the details.
How do remote and cloud tasks compare?
Codex Cloud runs tasks on OpenAI's machines from a published cloud environment with your GitHub repositories, dependencies and access settings. It needs ChatGPT sign in (Plus or higher) and does not support GitLab. The Codex Cloud guide walks through it, and Codex cloud environments covers setup in depth.
Kiro Web is, in its docs' words, "a browser-based development agent that helps you build software through conversation." It offers a default mode, Spec mode and an autonomous mode that delegates "an outcome from planning through implementation and a pull request or merge request." It works with GitHub or GitLab, requires a Pro, Pro+, Pro Max or Power plan, and "is available in US East (N. Virginia) us-east-1 only." The docs add: "Kiro never merges changes automatically."
How much do Codex and Kiro cost?
As of October 2026, the Codex pricing page lists ChatGPT Free ($0) and Go ($8 a month) with GPT-6 Luna in the desktop app, subject to rollout; Plus at $20 a month with Codex on the web, CLI, IDE extension and iOS; Pro from $100 a month; and Business at $20 per user a month billed annually. An API key works too, billed at API rates, with no cloud features.
As of October 2026, Kiro's pricing page lists:
| Plan | Price | Credits a month |
|---|---|---|
| Free | $0 | 50 |
| Pro | $20 per user a month | 1,000 |
| Pro+ | $40 per user a month | 2,000 |
| Pro Max | $100 per user a month | 5,000 |
| Power | $200 per user a month | 10,000 |
| Enterprise | Contact sales | Not listed |
Extra Kiro credits cost $0.04 each, and monthly credits do not roll over. Codex usage counts against your ChatGPT plan's usage limits, which you can extend with ChatGPT credits.
Which should you pick?
Codex tends to fit when:
- You already pay for ChatGPT, or prefer an API key with usage billing.
- You work mostly in the terminal or your current editor and want an open source CLI.
- You want a sandbox with network off by default and minimal setup.
- You want to send GitHub tasks to the cloud and pull diffs back with
codex apply.
Kiro tends to fit when:
- You want requirements, a design and a task list reviewed before any code is written.
- You prefer a full IDE based on Code OSS, with your VS Code settings carried over.
- You want to switch between Claude, GPT and open weight models under one plan.
- Your repositories live on GitLab as well as GitHub and you want remote tasks for both.
Many teams use both: specs in Kiro for large features, Codex for quick fixes and scripted runs. A shared AGENTS.md keeps them aligned, and the spec-driven development skill brings a Kiro-like flow to Codex.
Keep your prompts and skills portable
Skills are the part of your setup worth carrying between agents. Kipt keeps prompts and Agent Skills as Markdown in a private GitHub repo you own, imports skills from a GitHub link, a folder or a .zip, and installs every skill into Claude Code or Codex with one npx skills add command. For Kiro, copy the same folders into .kiro/skills/. The free SKILL.md generator turns a prompt you keep pasting into a skill folder.