CLI changelog

This release removes friction from MCP iteration and gives custom agent authors tighter control over context. Agent and MCP configs now reconcile live when you save a change, without restarting the session or losing conversation history. A new setting also lets custom agents opt out of inheriting default resources.

Agent and MCP configurations now hot-reload when you save changes on disk. Edit an agent config, add a new agent file, or modify mcp.json and the changes take effect without restarting the session. Only affected servers restart, and your conversation context is preserved. The config diff is order-independent, so reordering environment variables won't trigger unnecessary restarts. Learn more ->

The new chat.disableInheritingDefaultResources setting stops custom agents from inheriting default steering, skills, and AGENTS.md. Set it to true when you run only custom agents and don't want global resources pulled in. Since v2.7.0, custom agents inherit default resources automatically. This setting gives you a way to turn that off. Learn more ->

This release fixes session expiration for Entra ID (Azure AD) users, resolves approval prompt loops for compound shell commands in V3, and adds compact tool card previews for sub-agent calls.

This release introduces an early access preview of Kiro CLI V3. The new harness runs alongside your existing 2.x install, so you can try it without disrupting your current setup. Run kiro-cli --v3 to opt in.

Kiro CLI 3.0 is built on the same unified agent harness that powers the Kiro IDE and Kiro Web. Every improvement to the harness now ships to all surfaces simultaneously. V3 introduces spec-driven development in the terminal, a capability-based permissions model, enhanced hooks with a standalone file format, and tag-based agent configuration. Your existing 2.x setup is unchanged until you opt in with kiro-cli --v3. Learn more ->

This release gives you more control over the agent's direction and autonomy. Queue steering lets you redirect the agent at tool boundaries without waiting for it to finish, /goal creates iterative loops that verify completion before stopping, and the /rewind picker now shows exactly what happened at each turn.

Start a goal-driven loop where the agent works iteratively toward an objective and must verify the task is complete before finishing. Unlike a regular prompt, /goal keeps the agent running through multiple cycles of implementation and self-check (default: 5 iterations, configurable with --max) until the acceptance criteria are met. Useful for multi-step tasks where you want autonomous execution with a built-in quality gate. Learn more ->

Send messages to the agent while it's working, and it picks them up at the next tool boundary. Instead of cancelling and restarting when you spot a wrong turn, queue a correction and the agent adjusts course mid-turn. Press Ctrl+S to toggle between steer mode (inject mid-turn) and queue mode (buffer until turn ends). Learn more ->

The /rewind turn picker now shows a high-level summary of each turn — tool calls, files touched, commands run, and any intermediate messages so you can quickly identify the right branch point without expanding individual messages. Learn more ->

This release focuses on portability and getting back to work faster. Export any conversation to share or archive, label terminal windows so you know which session is where, and start a session at the reasoning level you need. Your model and effort choices now stick around on their own, so there's no extra step to make a preference the default.

Save a conversation to share with a teammate, attach to a ticket, or feed into another tool. Run /transcript save to export the full chat as markdown, plaintext, or JSON. The --plain and --json flags also switch the rendered transcript format inline, so you can read it the way you want without leaving the pager. Learn more ->

Keep track of which session is running where. Use /title to set or clear the terminal window title so each window has a meaningful label. Turn it on from /settings display → Terminal title. Learn more ->

Start a session at the reasoning depth the task calls for. The new --effort flag for kiro-cli chat sets the initial effort level (low, medium, high, xhigh, max) when you launch, so quick lookups stay fast and complex work gets the deeper reasoning it needs from the first prompt. Learn more ->

/model and /effort now remember your choices automatically. Switch models or change effort once and Kiro carries that preference into future sessions, so you no longer need to run /model set-current-as-default to make a selection stick. Learn more ->

This release makes Kiro's work more transparent and more autonomous. Watch the agent reason in real time with thinking display, let subagent pipelines review and revise their own work before handing it back, and control how output renders with new display settings. Message and shell output rendering are also noticeably faster during streaming.

See how the agent works through a problem as it happens. Thinking display streams the model's reasoning in real time, so you can follow its logic, catch a wrong turn early, and understand why it chose an approach. It's enabled by default. Toggle it from /settings > Display > Show thinking - the change applies to your next chat session. Learn more ->

Subagent pipelines can now revise their own work. A reviewer stage can send a task back to the implementer for changes and loop until the work meets the bar, all before results return to you. This turns multi-agent pipelines into self-correcting workflows for cases like code review and refactoring. Learn more ->

Run /settings display to control how Kiro renders output. Toggle streaming animations, decorative ASCII art, and icons versus text-only indicators to match your preference or accessibility needs. Changes take effect immediately, with no restart required. Learn more ->

This release gives you finer control over how conversations flow and how the model thinks. Rewind to any earlier prompt to explore a different path, set reasoning effort per model, and manage all your preferences from one unified settings menu. Workspace initialization is also 88% faster.

Jump back to any earlier prompt in a conversation and continue from that point in a new session. Use /rewind to pick a turn, then branch off without losing your original thread. This is useful when you want to explore an alternative approach or undo a direction the agent took without starting over from scratch. Learn more ->

Control how much thinking the model does with /effort. Choose from five levels: low, medium, high, xhigh, and max. Lower effort means faster, cheaper responses for straightforward tasks; higher effort gives the model more room to reason through complex problems. Learn more ->

Access theme customization, keyboard shortcuts, and terminal configuration from a single /settings command. Run /settings theme to change colors, /settings keybindings to view shortcuts, or /settings terminal to enable Shift+Enter / Option+Enter for multi-line input. The existing /theme command still works as a shortcut. Learn more ->

This release gives you more control over which MCP servers you can connect to, where Kiro lives, and how the TUI responds to your keystrokes.

Connect to HTTP-based MCP servers that don't support Dynamic Client Registration by setting a pre-registered oauth.clientId in your MCP config. This unlocks services like Slack, GitHub, and Figma that require manual app registration for OAuth, so you can use their MCP servers without running your own proxy. Learn more ->

Point Kiro at a different directory for global agents, prompts, skills, steering, settings, and sessions by setting KIRO_HOME before launching. This is handy when you manage dotfiles across machines, keep work and personal profiles separate, or need to isolate Kiro's state for containerized environments. Learn more ->

Remap the cancel, close-menu, and quit actions in the V2 TUI to keys that fit your muscle memory. If Ctrl+C and Esc conflict with your tmux prefix or terminal bindings, you can now pick different keys instead of working around the defaults. Learn more ->

Shell commands the agent runs can now stream output through two additional channels. Write to $AGENT_DISPLAY_OUT to show rich progress in the TUI without bloating the agent's context, and write to $AGENT_CONTEXT_OUT to surface lines in the tool result's agent_notes field. Stdout keeps working as before. Learn more ->

This release adds adaptive thinking support so model reasoning carries across turns in a conversation, giving you more coherent answers as context builds. It also fixes a silent failure in subagent tool dispatch when MCP servers update their tool specs mid-execution.

Kiro CLI now preserves model reasoning across multi-turn conversations. Thought content from earlier turns stays available to the model on subsequent turns, so follow-up answers stay aligned with the reasoning that produced previous responses. You get more coherent multi-step workflows without re-explaining context.

This release focuses on keeping you in flow: shell output streams as it happens, Tool Search loads MCP tools on demand so your context window stays clear, and skills are now invokable as slash commands. This release also adds device flow authentication for remote environments, RHEL support, and several rendering and stability fixes.

Shell command output now streams to the terminal line by line as it runs, instead of buffering until the process completes. Long-running commands like builds, test suites, and deployments show progress immediately, so you can follow along and catch issues early. Learn more ->

Tool Search loads MCP tools on demand instead of sending every definition with each request, keeping your context window clear when you have many MCP servers configured. Enable it with kiro-cli settings toolSearch.enabled true. Learn more ->

Skills defined in .kiro/skills/ are now available as /skill-name slash commands. Type / and the skill name to invoke it directly, giving you quick access to reusable agent instructions without switching agents or copying instructions into your prompt. Learn more ->

Sign in with Google or GitHub from SSH sessions, containers, and cloud workspaces without port forwarding. The CLI displays a URL and one-time code you enter in any browser — same flow as Builder ID and IAM Identity Center. Learn more ->

Kiro CLI's terminal UI now runs on Red Hat Enterprise Linux (RHEL). If you're working in enterprise Linux environments, you get the same full-featured terminal UI experience available on other Linux distributions, macOS, and Windows. Learn more ->