Kiro Studio

A graphical VS Code front end for the locally installed Kiro CLI, built on the Agent Client Protocol (ACP): a chat with streaming answers, diffs, checkpoints you can revert, and views for your sessions, specs, steering files, hooks, MCP servers and agents.

Unofficial. Kiro Studio is an independent project. It is not affiliated with, endorsed by, or sponsored by AWS or Kiro. "Kiro" is a trademark of its owner. This project ships no Kiro logos or other Kiro assets.

Preview. All phases of the plan are built and tested against a scripted stand-in for the CLI and, for the parts that do not need it, against real Git and a real VS Code. Apart from the author's probe runs it has not been tried against a real Kiro CLI, so expect rough edges. docs/PLAN.md lists what is assumed and how to check it.

Sample data, rendered by the project's own preview script (npm run preview) in three themes; not a recording of a real Kiro session.

What it does

• Chat in the sidebar or an editor tab, with sessions as tabs. Streaming Markdown, highlighted code, Mermaid diagrams, Kiro's reasoning as it streams (when the CLI sends it), "Kiro is thinking…" until a turn shows something, tool calls as cards with diffs, the plan, / commands, # context (files, folders, the open file, problems, git diff, steering files, specs) and pasted or dropped images. Agent, model and effort are picked in the input box; Supervised/Autopilot, the context fill and the credits of the last turn sit below it, and a chip above it opens Kiro's latest changes.
• Questions from Kiro are answered in the chat with exactly the options Kiro offered, with the change shown. In Autopilot, Kiro Studio allows each request once for you (see Safety); Supervised always asks.
• Checkpoints: every turn that changes files leaves a card with View changes and Revert, backed by a separate Git repository in the extension's storage. Your own repository is never touched. Needs Git.
• Kiro panel: your earlier sessions (open, search, fork where the CLI allows it), and your specs, steering files, hooks, MCP servers and custom agents, read from .kiro and ~/.kiro, with the actions to create, switch on and off, and remove them. Edits keep your comments and formatting.
• Specs: an assistant starts a spec (feature with requirements or design first, bug fix, quick spec) in the right Kiro agent. A view per spec shows requirements, design (diagrams drawn) and tasks with progress, and lets you approve a document or ask for changes. Start task over every open task in tasks.md asks Kiro to do it.
• Export a session as Markdown. English and German. Light, dark and high-contrast themes; keyboard and screen reader support; long sessions stay responsive.

Requirements

• Kiro CLI with the V3 engine, signed in (kiro-cli login). Run kiro-cli update if yours is older. The headless mode that needs KIRO_API_KEY is deliberately not used.
• VS Code 1.93 or newer. Windows 11, macOS and Linux. A real file system (no virtual workspaces) and a trusted workspace.
• Git for checkpoints (everything else works without it).

Getting started

1. Install Kiro Studio from the Extensions view (search for "Kiro Studio"), or a .vsix from the releases with Extensions: Install from VSIX…, and open a folder. The Kiro Studio icon appears in the activity bar.
2. If the CLI is not found or you are not signed in, the view says so and offers the fix. Kiro Studio: Open Getting Started walks through it.
3. Start a session with + in the chat and ask something. The first start of the CLI can take a while (up to about 20 seconds on Windows).

Settings

Privacy and safety

• Kiro Studio starts your local Kiro CLI and talks to it over standard input and output. It has no telemetry and makes no network requests of its own; the CLI talks to Kiro's service with your account, as it does in a terminal.
• It starts an agent that can read and change files and run commands, so it stays off in untrusted workspaces.
• Autopilot (with the setting above) answers allow once for you. Kiro's own deny rules still apply, because Kiro does not ask for what it denies; Kiro's ask rules are answered. A permanent "always" is never chosen for you, the end-of-turn review stays yours, and every automatic answer is shown in the chat. Switch the setting off to be asked every time.
• Hooks run only after you turn them on, and then also your global hooks from ~/.kiro/hooks.
• The protocol log hides tokens, e-mail addresses and your home folder. Values of env and headers of MCP servers are never read out; only their names are shown. Kiro's own session files are never read or written directly.
• Checkpoints live in the extension's storage, not in your repository.

Known limits

• Not yet checked against a real CLI beyond the author's probe runs; docs/PLAN.md (sections 9 to 12) lists every assumption, for example the format of tasks.md and what the CLI does when you reject a review.
• The scripted stand-in for the CLI is a shell script, so the integration tests that use it are skipped on Windows.
• Only the first workspace folder is tracked by checkpoints. Sessions of a remote Kiro are listed but not driven specially.

More

• docs/PLAN.md: architecture decisions, phases, risks, assumptions and how to check them (German)
• docs/ARCHITECTURE.md: how the pieces fit together
• docs/protocol-notes.md: what kiro-cli acp does on the wire, with the evidence for each statement
• CHANGELOG.md
• scripts/probe-acp.ts: starts the ACP server in a throw-away folder and records every message

Development

Requires Node 22 (the packaging tool needs it) and, to try the extension, a signed-in Kiro CLI.

npm ci
npm run build              # bundle to dist/extension.js; press F5 ("Run Extension") to try it
npm run typecheck && npm run lint && npm test
npm run test:integration   # downloads VS Code; the CI runs it on Linux, macOS and Windows
npm run package            # dist/kiro-studio.vsix

Releases: docs/RELEASING.md (one-time setup, the check with a real CLI, publishing; German).

Running the probe

Requires Node 20+ and a Kiro CLI that is installed and signed in (kiro-cli whoami).

npm ci
npm run probe -- --help
npm run probe -- --no-prompt    # handshake and session only, no credits
npm run probe                   # also sends one tiny prompt

Output goes to probe-logs/<timestamp>/ (summary.json, wire.ndjson, stderr.log), which is git-ignored. The logs are redacted on a best-effort basis; read them before sharing. npm run validate-wire -- <wire.ndjson> checks which recorded messages the ACP SDK's own schemas would reject.

License

Apache License 2.0, see LICENSE.