Kooky
A minimal modern terminal for AI coding: sidebar workspaces, split panes, one-click agent launch, live activity readouts, and one-click Node/branch switching.
README asserts multiple security behaviors: paste confirmation for suspicious multi-line commands, OSC 52 clipboard approval, secure keyboard input during sudo/ssh, no telemetry, kooky:// deep links cannot carry commands, CLI over a local owner-only socket. But evidence is only README and LICENSE — no source code corroborates any of it, so all are unsupported assertions scored 1. least_privilege: broad capability surface (SSH sessions, CLI tab control) with no code-level permission design visible. rollback: state persistence and session resume are described, but no update/rollback path. source_attribution: NOTICE.md and upstream theme licenses are referenced, but those files are absent.
failure_messages scores 2: CLI failure handling is best documented ('any failure prints one reason line on stderr', status exits 1 when not running). self_consistency: narrative is internally coherent, but claims like '1000+ unit tests' cannot be verified within the evidence — deducted. dependency_availability: depends on Xcode 26+ and a libghostty xcframework via a setup script whose existence is unverifiable.
Audience is clear (AI-coding terminal, macOS 14+), shell coverage (zsh/bash/fish) and multi-agent support are described concretely — audience and environment_fit get 2. capability_boundaries: the table explicitly marks unsupported per-agent capabilities (e.g., Grok lacks the waiting dot) — 2. trigger_precision: 'Ask agent', ⌘P and other triggers are text-only descriptions; implementation precision unverifiable.
license earns full marks: LICENSE is complete MIT text, README adds NOTICE.md and a badge. install_notes gets 2: three Gatekeeper bypass paths are thorough and honestly flag the adhoc-signature risk. examples gets 2: full CLI command examples. information_architecture and naming_stability get 2. known_limitations only 1: scattered mentions (agents lacking signals, SSH history local-only) but no systematic limitations section. versioning_changelog 1: CHANGELOG.md referenced but not provided. maintenance_responsibility 1: single unknown author, no maintenance commitment stated.
output_usability 2: CLI JSON output, status bar, notification center are well described for usability. marginal_value 2: agent integration (one-click launch, session resume, activity readout) is clearly differentiated from a plain terminal. cost_benefit only 1: no code or test evidence supports the claimed scale, so the benefit judgment rests on unverifiable assertions.
claim_traceability 1: many specific claims (fifteen agents, 1000+ tests, 40+ themes) cannot be traced to provided files. cross_source_corroboration 1: CN/JA READMEs, CHANGELOG, NOTICE are referenced but absent from evidence. fact_inference_separation 2: the README separates 'what it does' from 'how to configure' cleanly, without heavy promotional conjecture.
- This review is based solely on README and LICENSE — no source code, tests, or CHANGELOG were provided; all security and reliability claims (paste confirmation, OSC 52 approval, owner-only socket, 1000+ tests) are unverified.
- The app is adhoc-signed and requires a Gatekeeper bypass on first launch; README Path C disables Gatekeeper system-wide, a system-level risk — use only when necessary and re-enable immediately.
- Publisher identity is unverified; verify binary provenance before installing prebuilt .dmg files, and in enterprise settings audit or build from source.
- kooky-cli can open/focus/close tabs and launch agent sessions from any local script; evaluate the local socket's permission boundary before deployment.
What does this agent do, and when should you use it?
Kooky is a SwiftUI-native macOS terminal purpose-built for running AI coding agents, rendering through the GPU-accelerated libghostty engine. It ships sidebar workspaces, horizontal/vertical split panes, and one-click launch for fifteen AI agent CLIs (Claude Code, Codex, Gemini CLI, and more), with per-agent activity readouts including waiting dots and tool pills. All fifteen built-in agents resume their conversations across kooky restarts using each CLI's exact session ID, and a right-side agent panel aggregates agents across every window sorted by who needs you first. It also covers Git worktree creation, remote SSH workspaces, a sidebar file tree, a multi-line prompt composer (⌘L), Quick Open (⌘P), and a local CLI named kooky-cli for scripting. The app has no accounts, no telemetry, and no cloud sync — state stays local — and is MIT-licensed open source.
Kooky opens folders as workspaces; each workspace splits into panes with independent tab strips and sessions. From the '+' menu you launch one of fifteen agent CLIs (claude, codex, gemini, opencode, amp, cursor-agent, copilot, grok, agy, kimi, pi, omp, reasonix, kiro-cli, droid), or add a custom agent in Settings → Agents (own logo, environment variables, optional built-in template inheritance). Sidebar dots track agent state live — running (blue), waiting on you (amber), idle (none) — while the pane status bar shows the currently running tool, git repo/branch/diff, Python venv, Node version, and active proxy; clicking the Node or branch pill switches without typing. Right-click any git workspace → 'Create Worktree…' for a nested worktree with its own tabs and agent; ⌘P opens SSH workspaces where agents start on the remote with the remote's shell setup, and pasted local files upload first. Select text → right-click → 'Ask <agent>' submits the selection as the first prompt; ⌘L opens a multi-line prompt composer; the agent panel's history page browses and resumes past conversations by exact session ID; Session Info shows process trees, listening ports, exit codes, and an 'Ask AI' button for failed commands. kooky-cli opens tabs, resumes conversations, and lists/focuses/closes tabs from scripts.
- Developers running Claude Code or Codex across multiple projects in parallel: sidebar workspaces plus split panes keep each project separate while the agent panel surfaces whoever is waiting first.
- Frontend engineers switching Node versions or git branches constantly: click the status-bar pill to switch — no typing.
- Engineers who want Claude working on a feature branch without touching main: right-click workspace → Create Worktree… spawns a nested worktree.
- Remote-development users with nvm-installed tools and remote shell setups: SSH workspaces launch agents on the remote machine, and pasted local files upload automatically.
- Automation users scripting their terminal: kooky-cli open/resume/list/focus/close drives kooky from other local apps.
- Anyone pasting commands from web pages: multi-line pastes outside bracketed-paste framing require confirmation, blocking booby-trapped curl … | sh from executing on ⌘V.
What are this agent's strengths and limitations?
- Native integration with fifteen mainstream AI coding agent CLIs — one-click launch and cross-restart session resume for all of them via each CLI's exact session ID.
- GPU-accelerated rendering via libghostty stays smooth and tear-free on 120Hz/ProMotion displays; SwiftUI-native UI with full IME support.
- Rich live status: sidebar dots, tool pills, cross-window notification inbox, and an optional menu-bar agent glance — waiting agents notify you even in unfocused tabs.
- SSH workspaces start agents on the remote with the remote's own shell setup, and pasted local files upload and become remote-openable paths.
- Three-state Keep-awake: the Mac stays awake (lid closed included) while an agent works or an SSH session is live, and sleeps the moment work ends.
- macOS 14+ only — no Linux or Windows builds.
- Ad hoc signing means manual Gatekeeper bypass on first launch; no Apple Developer ID public-distribution signing or notarization yet.
- Building from source requires Xcode 26+ plus a separate script to fetch the libghostty xcframework — a nontrivial build setup.
- Waiting dots are unavailable for Grok Build and Kiro CLI (no signal exposed), and the tool pill is not supported by every agent.
- Some features (secure keyboard input during sudo/ssh, lid-closed keep-awake) require one-time macOS admin authorization, and remote programs' clipboard access over OSC 52 needs per-request approval.
How do you install or deploy this agent?
Download the latest .dmg from GitHub Releases and drag Kooky.app into Applications. First launch is blocked by Gatekeeper (ad hoc signing): Path A — System Settings → Privacy & Security → Security → Open Anyway; Path B — run xattr -d com.apple.quarantine /Applications/Kooky.app. Building from source requires Xcode 26+ and macOS 14+: run ./scripts/setup-libghostty.sh (one-time xcframework fetch), then swift build, then ./scripts/build-app.sh or ./scripts/build-dmg.sh --build.
How do you use this agent?
Open Kooky, create a workspace with ⌘O or by dragging a folder onto the sidebar; click '+' to launch an agent (e.g. Claude Code) in one click; ⌘D / ⌘⇧D to split; ⌘P for Quick Open or new SSH workspace; ⌘L for the multi-line prompt composer; right-click a workspace for worktree creation and color tags; the agent panel (right sidebar) shows, jumps to, and resumes every agent including past conversations; select an error and right-click 'Ask <agent>' to submit it directly. CLI examples: kooky-cli open --cwd ~/Github/vibex -e "npx @deepseek-ai/dsh web", kooky-cli resume --agent codex --id <conversation-id>, kooky-cli list --.
How does this agent compare with similar options?
The README repeatedly references Ghostty: kooky uses Ghostty's libghostty rendering engine, inherits Ghostty config/themes, and mirrors ghostty.app's click-to-position-cursor UX — positioning it as an AI-agent-workflow terminal built on top of the Ghostty engine rather than a general-purpose terminal emulator.
FAQ
Is there a cost or account requirement?
First launch is blocked — what do I do?
xattr -d com.apple.quarantine /Applications/Kooky.app. macOS only blocks the first launch.