OctoShell banner
savvasioakeim savvasioakeim

OctoShell

AI community

Description

Mission control for coding agents on Windows. An AI orchestrator fans tasks out to multiple coding agents, each in its own git worktree, from one terminal workspace.

Installation

This entry records only its repository, not the path inside it, so there is no exact command to give. Open the source below and copy the folder into ~/.claude/skills/, or the file into ~/.claude/agents/.

README

🐙 OctoShell

**Mission control for coding agents — Windows-native.**

OctoShell is a power-user terminal workspace where an AI **orchestrator** fans your tickets out to multiple coding agents, each working in its **own isolated git worktree**, while you watch every terminal, agent trace, and dev server from one window.

Built with **Tauri v2 + Rust** (backend) and **React + TypeScript + Tailwind + xterm.js/WebGL** (frontend).

🇬🇷 Ελληνικά: [README.el.md](README.el.md)

What it does

  • Orchestrator sidebar — a workspace-wide AI assistant that sees every open project (terminal blocks, agent reports). Ask it to coordinate work: it proposes dispatch actions (confirm-per-click or fully autonomous), creates a git worktree per task, and follows each agent to completion with live watch.
  • Agents per project/worktree — Claude Code and Gemini CLI natively, plus any ACP agent (Agent Client Protocol: Claude, Gemini, Codex, OpenCode, Cursor, Copilot, Kiro) over one protocol integration. Per-project provider, model, profile and approval mode.
  • Review Mode — when dispatched features are done, a floating always-on-top window walks you through them one by one: start their dev servers with one click, take notes, approve/decline. Declines are batched back to the right worktree's agent as fix tasks.
  • Managed dev servers — server commands (npm run dev, vite, uvicorn…) run as managed services with their own port and logs instead of wedging a shell or an agent turn. Respects your project's .env PORT, validates npm scripts against package.json, and detects the actually-bound URL.
  • Docker sandbox (opt-in) — run an entire ACP agent inside a throwaway container (--cap-drop=ALL, memory/pids limits, worktree bind-mounted) with a shared login volume, so one login covers every worktree. Host isolation for every command the agent runs.
  • Semantic Blocks terminal — the terminal is a feed of self-contained command blocks (à la Warp), d