See what's done without reading any output.
Each card shows the task, the CLI, the branch and how long it has been in its state. Ready cards add a diff stat of what changed.
Every Claude Code, Codex, Cursor CLI, or Pi task gets its own git worktree and terminal. Shika tells you when one is ready.

Interactive prototype, sample data. Try jk to move, ↵ to focus the terminal, or click + to add a shell tab.
Prototype with sample data. Open it on a laptop to try the keyboard flow.
Open full screenShika is built for the keyboard. Your hands stay put from the first prompt to the push.
Pick Claude Code, Codex, Cursor CLI, or Pi. Shika creates a worktree and starts the CLI inside it.
Type straight into the real terminal. The first line names the branch, then the CLI's own title renames it.
When the agent goes quiet, a notification tells you which task is ready to check.
Add a shell tab in the same worktree. Diff, test, commit and push yourself.
Shika removes the worktree, and asks first if any work would be lost.
Each card shows the task, the CLI, the branch and how long it has been in its state. Ready cards add a diff stat of what changed.
Native macOS banners, titled with the project and the task. Click one to land on that card. One per turn, not one per redraw.
Agents never touch your main checkout or each other. Worktrees live in .worktrees/, kept out of git by the repo's local exclude, not your .gitignore.
~/code/routekit ├─ .worktrees/ │ ├─ fix-flaky-timeout-test │ ├─ json-route-listing │ └─ shika-draft-3f2a └─ src/ untouched
Uncommitted changes or unpushed commits? Shika tells you, then lets you discard or push. It never commits for you.
Built on GPUI, the GPU-accelerated UI framework behind Zed. Each agent gets a real terminal that keeps running while hidden, so a busy CLI never stalls.
Early-stage, macOS-only software. Build a local app from source; signed and notarized downloads are not available yet.
$ git clone https://github.com/xuanhieu2611/shika.git $ cd shika $ ./scripts/bundle-app.sh $ open target/release/Shika.app
Local bundles are ad hoc signed. Read the agent permission model before running Shika. Move Shika.app to Applications to keep it in the Dock.
Something else? Open an issue.
Yes. It is MIT licensed, with no account and no paid tier. You pay your agent's provider the way you already do.
Claude Code, Codex, Cursor CLI, and Pi today. Kiro is next. Shika starts the binary you already have, so your login, settings and plan come with it.
Shika's presets use the CLIs' automatic approval modes. Claude Code, Codex, and Cursor CLI bypass permission prompts or sandboxing; Pi starts with --approve. Agents can execute commands with your user account's access. A worktree separates Git edits but does not restrict access to files, credentials, or the network. Run agents only in projects you trust. Shika separately asks for approval before running configured worktree setup commands.
Shika has no telemetry or model API integration. Your agent CLI talks to its provider, Git fetch and push connect to your remotes, and any worktree setup commands you approve can use the network. Your CLI controls what it sends to its provider.
So agents can edit the same repo at the same time without colliding, and your main checkout stays clean. Each worktree is a full checkout on its own branch, under .worktrees/ in your repo.
It never commits. Add a shell tab with ⌘T, then review, commit and push yourself. When you close a card with a clean tree and unpushed commits, Shika offers to push them. A push keeps the card open until you close it.
Your projects and worktrees stay on disk. Running sessions end, and the next launch lists leftover worktrees so you can clean them up when you choose.
Not yet. Shika is a native Mac app for now.
After Shikamaru, the one who keeps track of everyone on the field while giving directions. Shika also means deer in Japanese, hence the logo.

Free, open source, and native to your Mac.