Kiro is Amazon Web Services’ agentic coding harness: the Kiro CLI reads codebases, writes code and runs commands in the terminal, alongside the Kiro IDE and web and mobile interfaces. As an AI harness, Kiro relies on spec-driven development, steering files, hooks and a declarative permission system; the software is proprietary and provided under the AWS Intellectual Property License. As of 20 September 2026.
What is Kiro?
Kiro comes from AWS. Per its documentation, Kiro CLI is the next update of the Amazon Q Developer CLI: available since 17 November 2025, existing Q Developer workflows, subscriptions and sign-ins keep working, but new features now ship only in Kiro. The entry point is kiro-cli instead of q, and Amazon Q rules became steering files. Unlike the Apache-licensed Q CLI, Kiro CLI is licensed under the AWS Intellectual Property License; its kirodotdev/Kiro repository, per its README, serves only as an issue tracker.
The CLI installs with curl -fsSL https://cli.kiro.dev/install | bash and starts with kiro-cli inside a project directory; macOS, Linux and Windows are supported. The vendor describes the IDE, CLI, web and mobile surfaces as interfaces to the same agent harness, an independent process addressed through the open Agent Client Protocol (ACP) – so kiro-cli acp also lets you use Kiro inside JetBrains IDEs and Zed.
Kiro as an AI harness
The defining concept is spec-driven development: a spec consists of requirements.md (or bugfix.md), design.md and tasks.md, leading from a requirement through a design to individually trackable tasks. Project knowledge comes from steering files under .kiro/steering/ (globally ~/.kiro/steering/) with inclusion modes such as always, fileMatch and manual; Kiro additionally reads AGENTS.md at the project root and in subdirectories.
Hooks live as JSON files under .kiro/hooks/ and react to events such as PostFileSave or PreToolUse with a shell command or agent prompt. Custom agents (.kiro/agents/) define tools, context and permissions for recurring tasks and can also run as subagents in parallel with their own context window. Skills follow the Agent Skills format with SKILL.md; Powers bundle MCP tools with domain knowledge as installable Agent Plugins. MCP servers are listed in .kiro/settings/mcp.json. The permission system uses capabilities such as fs_write, shell or mcp with the effects deny, ask and allow in ~/.kiro/settings/permissions.yaml. Sessions save automatically and resume with kiro-cli chat --resume.
Models and model access
Kiro offers an Auto mode that routes every task to a suitable model, plus a curated selection: Claude models from Anthropic, GPT models from OpenAI, and open-weight models such as DeepSeek, MiniMax, GLM and Qwen. Usage is measured in credits with a model-dependent multiplier. The models are served through the Kiro service, which according to the documentation is built on Amazon Bedrock.
The documentation does not describe a freely configurable base URL or custom model endpoints. Whether model access can be routed through a central proxy such as innFactory’s AI Gateway is therefore not publicly documented; what is documented is network proxy settings and a list of domains to allow. Model control instead runs through approved-model lists in the Kiro console.
Enterprise use: data protection and governance
AWS distinguishes three user groups. Content from free-tier users and individual subscribers is stored in US East (N. Virginia) and, subject to an opt-out, may be used for service improvement; content from enterprise users is not used this way and instead stored in the region of the Kiro profile, with Europe (Frankfurt) supported alongside US East. Inference runs via cross-region inference within the geography – in Europe, for example, across Frankfurt, Ireland, Paris and Stockholm. GPT models, however, are served from the US regardless of profile region, and models marked experimental may be processed globally.
Administrators restrict models and MCP servers in the Kiro console, control API key generation, and disable web tools; sign-in works via AWS IAM Identity Center, Okta or Microsoft Entra ID, and customer-managed KMS keys are available. Headless mode (kiro-cli chat --no-interactive) uses KIRO_API_KEY under the same policies. In every case, source code and prompts go to the Kiro service and its model providers – map this in advance in your records of processing activities.
Orchestrating Kiro with herdr
The terminal orchestrator herdr lists Kiro CLI as a supported agent, reading its state – working, blocked on an approval, or waiting – from the terminal output via a screen manifest; herdr’s documentation names no deeper integration role for Kiro, which is enough for monitoring several Kiro sessions alongside other agents.
Where it fits: developer harness and enterprise harness
Kiro is a harness for developers and fits particularly well with teams already on AWS. For business users outside software development, CompanyGPT is the matching harness: tools via MCP servers, integrations and n8n, context via companyRAG, roles from Entra ID or Keycloak, and audit logs, run in a customer’s own Azure environment or, sovereignly, on STACKIT. The AI Gateway forms the shared cost and governance layer for harnesses with a freely configurable endpoint; see the overview of all AI harnesses for further coding agents such as Claude Code. We help organisations adopt Kiro and comparable harnesses with clear policies and the right region choice.