Agent Upstart

Cool things people are building with agents — and practical help getting the most out of them.

Playbooks · Coding tasks

Kiro CLI 3.0: V3 is the default, Classic is gone

Written November 7, 2026

Kiro CLI 3.0.0 makes V3 the default and retires Classic, the desktop app, and autocomplete. What the auto-upgrade changes, and what to check before it runs.

What 3.0.0 does

Kiro CLI 3.0.0, released November 3, 2026, makes V3 the default agent harness for interactive sessions on hosts that can run it. V3 is the same harness that powers Kiro IDE and Kiro Web, and you no longer need --v3 to get it. Existing V2 users are asked before the default applies, and Kiro upgrades custom agents still in the V2 format unless you turn that off. Kiro CLI is now a single standalone kiro-cli binary that retires the Classic interface, the desktop app, and inline autocomplete.

The retirements are specific. The Classic (V1) interactive experience no longer runs: if a command, alias, or setting still selects it, Kiro offers Continue with V3 when you start a chat, and scripts ignore the selection. To keep the V2 agent, run kiro-cli --v2 — the flag applies to that launch only. The desktop app no longer ships, and existing macOS desktop app installs switch to the standalone kiro-cli on their next update check. Inline shell autocomplete and inline suggestions are retired with the desktop app and have no standalone replacement; the autocomplete, inline, translate, theme, and dashboard commands are unavailable, and running kiro-cli doctor now directs you to kiro-cli diagnostic or kiro-cli issue.

The default applies differently by surface. New interactive sessions start on V3, and V2 users are asked first. Headless scripts using --trust-all-tools start on V3 by default; other scripts stay on V2. New ACP clients start on V3; existing clients stay on V2 unless already supported.

V3 brings capabilities V2 does not have. Cloud sessions let you disconnect and resume later from /sessions — start one with kiro-cli --cloud, attach a repository, and let Kiro work in a managed cloud sandbox. Memory, once enabled, carries repository context forward. Approve for me, once enabled, approves routine actions, blocks unsafe ones, and asks about the rest, while your deny, ask, and allow rules still decide first. /sandbox puts agent shell commands and their child processes inside an operating-system sandbox, and local stdio MCP servers can run inside the same sandbox. Workflows show their plan before starting. Workspace trust, added in 2.33.0, asks you to trust a folder before it loads that folder's hooks, agents, MCP servers, steering, workflows, or language server configuration — a cloned repository can no longer change how the agent behaves until you say so.

The migration

  1. Confirm the version, and keep the V2 escape hatch. Run kiro-cli --version and confirm it reports 3.0.0 or later. Keep kiro-cli --v2 available until your V3 checks pass: V3 can resume its own sessions and your V2 sessions, but V2 cannot open a V3 session, because the stored session formats differ.
  2. Answer the first-launch prompt. Run kiro-cli with no harness flag and answer the prompt, or run kiro-cli --v3 to try V3 in a new session. Choosing "Don't ask again" does not keep you on V2 — your setup is still upgraded to V3 when 3.0 arrives.
  3. Read the agent upgrade before it rewrites anything. The auto-upgrade runs when a V3 session starts unless you turned it off, and it rewrites each agent file in place after saving a .json.bak backup. To turn it off, run kiro-cli settings chat.enableAutoAgentUpgrade false. Before accepting the upgrade, diff the backup against the rewritten file: prompts, model selections, resources, and MCP server definitions carry over unchanged, and supported settings convert automatically, but anything the migration cannot convert is reported as a warning and needs a manual update. Run /upgrade-agent diagnostics after migration and address the conversion warnings before relying on the migrated rules. For a shared agent directory, run kiro-cli agent upgrade <dir> with --dry-run first to preview the changes.
  4. Fix permissions by hand where the model changed. V2 used trusted-tool lists and per-tool settings; V3 uses rules for actions and resources, and where rules overlap, deny takes precedence over ask, and ask over allow. V3 does not have the use_aws tool — AWS CLI commands run through the shell tool instead — so recreate use_aws allow and deny settings as shell permission rules, and keep them specific rather than allowing broad patterns such as aws *.
  5. Remove the retired commands from scripts. Delete autocomplete, inline, translate, theme, and dashboard from scripts and aliases, along with --classic and --legacy-ui. A script that still selects Classic is ignored silently, which reads as a script that stopped working for no reason.
  6. Answer the trust prompt per folder. The first time you start an interactive V3 session in a folder, Kiro shows a Trust this workspace? screen before it loads the folder's hooks, agents, MCP servers, steering, workflows, or language server configuration. The answer is saved per workspace root, so the screen appears once per folder — except for your home directory, where Kiro asks again at each launch.
  7. Try the new controls in one workspace first. Turn on Approve for me and /sandbox in a single project before enabling them everywhere. Approve for me is stored per workspace in the permissions file, so it stays with that repository.

A patched flaw worth knowing about

Kiro's history is the reason the workspace-trust and permission changes matter. In July 2026, Intezer disclosed a prompt-injection chain in which hidden instructions on an ordinary web page — one-pixel white text on an API documentation page — could make Kiro rewrite its own MCP server configuration (~/.kiro/settings/mcp.json) and run attacker code on the developer's machine, with no approval step able to stop it. The researcher only had to ask Kiro to summarize the page. AWS tracks the underlying file-write flaw as CVE-2026-10591 — insufficient access controls in the file-write tool in Kiro IDE before version 0.11, fixed in 0.11 — and says the same identifier covers the Intezer chain; Intezer confirmed the attack failed in 0.11.130. Kiro now marks mcp.json, .vscode/tasks.json, the .git directory, and other sensitive files as protected paths, each requiring explicit approval before a write. The flaw is patched, not current — and it is the reason a cloned repository's hooks and MCP servers now wait for your trust decision in V3.

What can go wrong

Limits

Related playbooks

Sources

Back to: Coding playbooks hub · Also: Google Jules ends Nov 11: export your work · What version should I be on · Vet an Antigravity plugin before the one-click install