Hydra — Multi-AI Agent Orchestrator

Desktop application that coordinates multiple AI agents (ChatGPT, Gemini, Claude) through real browser sessions controlled via CDP (Chrome DevTools Protocol) with Playwright, shared context, and local task management.

What it does

• Launches real browsers (Chrome / Edge) per agent via CDP
• Injects prompts and collects responses automatically
• Manages projects, knowledge base, and tasks locally (SQLite)
• Exposes a REST Context Server + MCP for CLI agents
• Electron dark-mode UI with 3-column layout

UI — current layout

┌─────────────┬──────────────────────┬───────────────┐
│  Left Rail  │     Workspace        │ Right Sidebar │
│  280px      │     (flex)           │   260px       │
│             │                      │               │
│  Projects   │  Active project      │  Browser      │
│  Agents     │  Task Broadcast      │  Sessions     │
│             │  Responses           │  (compact)    │
└─────────────┴──────────────────────┴───────────────┘

Notable UI features:

• New project button opens a modal with name, description, root folder (native Windows folder picker), and mode fields
• Browser Sessions appear in the right sidebar — each agent has an animated status dot (idle / working / done / error) and the current task visible
• Full dark mode with CSS variables

Architecture

Electron UI (Renderer — React)
  -> IPC (preload.cjs)
Electron Main
  -> ipcHandlers.js      (get-config, open-agent, inspect-agent, send-task, save-decisions, select-folder)
  -> playwrightManager.js
  -> toolBridge.js       (Hydra tool bridge)
  -> serverProcess.js    (starts Context Server)

Context Server (Express + SQLite + MCP)
  -> routes/projects, agents, tasks, sessions, settings, ai, context, todos
  -> db/schema.js + queries.js
  -> orchestrator/sessionRunner, executor, contextBuilder, safetyGuards

Project structure

agent-sync/
├── README.md
├── ARCHITECTURE.md
├── SETUP.md
├── ROADMAP.md
├── injectors/
│   ├── chatgpt.js
│   ├── claude.js
│   └── gemini.js
├── server/
│   ├── index.js
│   ├── mcp.js
│   ├── ai/
│   ├── db/
│   ├── orchestrator/
│   ├── prompts/
│   └── routes/
└── src/
    ├── main/
    │   ├── index.js
    │   ├── ipcHandlers.js      ← select-folder dialog added
    │   ├── playwrightManager.js
    │   ├── toolBridge.js
    │   ├── windowManager.js
    │   └── ...
    ├── preload/
    │   └── preload.cjs         ← selectFolder exposed
    └── renderer/
        ├── App.jsx              ← 3-column layout
        ├── styles.css           ← full dark mode
        └── components/
            ├── ProjectPanel.jsx ← New Project modal + folder picker
            ├── BrowserSessions.jsx ← compact sidebar with status
            ├── AgentSidebar.jsx
            ├── TaskBroadcast.jsx
            ├── ResponseCollector.jsx
            └── OrchestratorPanel.jsx

Quick Start

npm install
npm run db:init
npm run dev

Production build:

npm run build:renderer
npm run start

Workflow

1. Press New project → fill in name, description, select root folder from Windows, choose mode.
2. Add agents from the left rail.
3. In the right sidebar press Open browser per agent and log in manually.
4. Send tasks from TaskBroadcast — agents appear with status working and the current task visible.
5. Responses appear in ResponseCollector; save them to the knowledge base.

Documentation

File	Role
SETUP.md	Installation and configuration
ARCHITECTURE.md	General technical architecture
ROADMAP.md	MVP and next steps
APP-ERRORS.md	Centralized error log (auto-generated)

Monitoring & Task Workflow

Task Workflow (Retry & Templates)

• Retry: If a task fails, you can press the Retry button directly in the Orchestrator chat to resubmit the request.
• Templates: Use the Templates button in the Orchestrator info panel to load predefined prompts (e.g. Project Audit).
• Cleanup on startup: The app automatically cleans up "stuck" tasks (those left in the working state after a crash) and resets agents.

Error Monitoring

To monitor application health and collect errors in one place:

npm run monitor:errors

This script will generate or update the APP-ERRORS.md file with all recent errors from session logs.