KO | EN

GitLite

Minimalist Open Source Repository Search Engine

gitlite — search

#typescript #ai-agents #developer-tools #cli #codex #claude-code #dsh-plugin #mcp #react #agent #dsh #deepseek-harness

executive

★ 104 Open GitHub ↗

An app to maximize the very best of human developer skills: executive control of AI Agents

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Executive

> **This project was just open-sourced and is under active development.** It was built as an internal tool and is being adapted for public use. Expect rough edges. Please check the [Known Issues](#known-issues) section before installing, and [open an issue](../../issues) if you run into problems.

A real-time Claude Code orchestration dashboard for managing multiple AI coding sessions across machines.

Let humans focus on what humans are good at -- creativity, executive control, decision-making -- and let AIs focus on what they're good at -- lightspeed execution.

Built as an internal dev tool for [Vibe Otter](https://vibeotter.com) and open-sourced to help other vibe coders out.

[![Demo Video](https://img.youtube.com/vi/z-KV7Xdjuco/maxresdefault.jpg)](https://youtu.be/z-KV7Xdjuco)

## The Problem

If you're running 3-5 Claude Code sessions at once, you already know what happens: you kick off a task, wait for Claude to finish, get distracted scrolling X, and then come back and talk to whichever Claude finishes first -- even if it's working on something low-priority. Once there are more than two sessions running, your weakly human executive brain function simply cannot keep track of what's happening where.

Executive fixes this. It gives you a single dashboard that shows every Claude session across every machine, what it's working on, how important it is, and whether it's done. When something finishes, you hear a chime and know exactly what to work on next.

The goal is to get back to the kind of deep concentration you had when you artisanally coded by hand -- 4-6 hour sessions of intense focus -- except now you're pushing out multiple features in a few hours instead of a few functions.

## The Workflow

This tool was built around a specific workflow for managing Claude Code at scale:

- **Development server** (1-3 Claudes): A complete replica of production where Claudes run wild on new features. If they break something, roll back and try again.
- **Production server** (1-2 Claudes): Tech support, analytics, manual admin tasks that haven't been automated yet. One Claude might be analyzing user pain points while another helps a customer.
- **Local machine** (1 Claude): Marketing tasks, side projects, LinkedIn posts, whatever doesn't need a server.

The standard process for each task:
1. Plan in detail with Claude -- discuss and refine until you're on the same page
2. Approve the plan
3. Turn on autopilot and don't talk to it again until it's done
4. Move on to the next Claude that needs attention

## Features

### Real-Time Task Dashboard

Every Claude session auto-registers when it starts. You see live status updates, working/done/queued states, and hear a chime when something completes. No more checking terminals to see if Claude is still thinking.

### Priority Tiers

Tag tasks as **routine**, **important**, or **urgent**. When multiple Claudes finish at once, you know which one to deal with first.

### Autopilot Mode

The killer feature. Once you've thoroughly planned and discussed Claude's approach, flip the autopilot switch. Executive will auto-approve all tool calls and permission requests, letting Claude work for 20+ minutes without interruption. This is a massive productivity unlock -- you can fully context-switch to another task for real.

> **WARNING: Autopilot mode auto-approves ALL tool calls and permission requests without human review.** This increases susceptibility to prompt injection attacks. If a repository or file Claude is working with contains malicious instructions, autopilot will execute them without asking. **We recommend NOT using autopilot when working with repositories from unknown or untrusted sources** -- which, ironically, includes this one if you just found it on the internet. Use at your own risk.

### Multi-Machine Support

Two deployment variants:
- **`local/`** -- Runs on `localhost:7777`, trusts localhost connections, optional API key for remote access. Best for single-machine use.
- **`cloud/`** -- Runs on `localhost:7778` behind a reverse proxy (nginx), password auth with bcrypt + signed HTTP-only cookies. Best for multi-machine setups accessed over the internet.

## Quick Start

> **Platform support:** Linux and macOS only. Windows is not yet supported -- see [Known Issues](#known-issues).

### Local (Single Machine)

```bash
cd local
npm install
./setup-machine.sh
./start.sh
```

Open `http://localhost:7777` in your browser. Start a new Claude Code session -- it will auto-register with the dashboard.

### Cloud (Multi-Machine)

On the server:

```bash
cd cloud
npm install
./setup-machine.sh
./start.sh
```

Set up nginx as a reverse proxy to `localhost:7778` with HTTPS. Then on each remote machine:

```bash
./setup-machine.sh machine-name https://executive.yourdomain.com
```

You'll be prompted for the API key from the server's `.env` file.

## How It Works

Executive integrates with Claude Code through [hooks](https://docs.anthropic.com/en/docs/claude-code/hooks) -- shell scripts that run at key lifecycle events:

| Hook | Event | What It Does |
|---|---|---|
| `session-start.sh` | SessionStart | Registers the session with the dashboard |
| `autopilot-hook.sh` | PreToolUse | Checks autopilot status, auto-approves if enabled |
| `permission-hook.sh` | PermissionRequest | Auto-approves permission dialogs when autopilot is on |
| `resume-hook.sh` | UserPromptSubmit | Marks task as "working" when you send a new message |
| `stop-hook.sh` | Stop | Marks task as "done" when Claude stops |
| `session-end.sh` | SessionEnd | Removes the task from the dashboard |

The dashboard uses Server-Sent Events (SSE) for real-time browser updates -- no polling, no WebSockets, no dependencies.

## Configuration

### Environment Variables (Cloud)

| Variable | Description |
|---|---|
| `EXECUTIVE_PASSWORD_HASH` | bcrypt hash of your dashboard password |
| `EXECUTIVE_API_KEY` | API key for hook-to-server authentication |
| `EXECUTIVE_COOKIE_SECRET` | Secret for signing session cookies |
| `EXECUTIVE_PORT` | Server port (default: `7778`) |

These are generated automatically by `setup.js`. See `cloud/.env.example` for the template.

### Home Directory Files

The setup script writes these files for the hooks to use:

| File | Purpose |
|---|---|
| `~/.executive-key` | API key for authenticating with the dashboard |
| `~/.executive-machine` | Machine name displayed in the dashboard |
| `~/.executive-host` | Dashboard URL that hooks call |

## Known Issues

### Windows is not supported (fix targeted for Feb 6, 2026)

Executive currently only works on **Linux and macOS**. The hooks are bash scripts that use Linux/macOS-specific commands (`lsof`, `readlink -f`, `curl`), and the setup scripts assume a Unix environment. Running on Windows -- including Git Bash -- will not work. WSL (Windows Subsystem for Linux) may work but is untested.

Windows support is actively being worked on and is expected by **February 6, 2026**.

### Setup may overwrite your Claude Code settings

The `setup-machine.sh` script configures Claude Code hooks by writing to `~/.claude/settings.json`. **It completely replaces the `hooks` key**, which means any existing hooks you had configured will be overwritten. The script creates a `.bak` backup of your settings file before writing, but if your `settings.json` fails to parse (e.g. it contains comments or is malformed), the backup is skipped and your settings will be overwritten without recovery.

**Before running setup**, manually back up your settings:

```bash
cp ~/.claude/settings.json ~/.claude/settings.json.backup
```

If you already ran setup and lost settings, check for `~/.claude/settings.json.bak`.

## About

Executive was built as an internal development tool for [Vibe Otter](https://vibeotter.com), an AI website builder for small businesses. The Vibe Otter codebase is 99.999% vibe-coded with Claude Code (6 lines were written by hand, last June) and has built over 1,000 websites for customers. Executive is the tool that made it possible to orchestrate multiple Claude sessions across development, production, and local machines simultaneously.

Open-sourced to help other vibe coders manage their AI coding workflows.

## License

executive

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README