Dorothy: AI Agent Orchestration Platform

Free, open-source (MIT) desktop app for orchestrating multiple AI coding agents simultaneously. Built by Charlie85270. Tagline: "Your AI Agents, Perfectly Managed."

Architecture (Three Layers)

1. Electron Desktop Shell

• Renderer: React 19 / Next.js 16 (App Router) — agent dashboards, Kanban, automations, scheduled tasks, usage stats, skills, plugins, settings
• Main Process: Agent Manager (node-pty-based), PTY Manager, background services (Telegram bot, Slack bot, Kanban automation, MCP server launcher, API server)
• API Routes: Next.js backend; renderer↔main IPC communication

2. Five MCP Servers (stdio-based)

• mcp-orchestrator (26+ tools): Agent lifecycle, messaging, scheduling, automation management, JIRA updates
• mcp-telegram (4 tools): Messaging with photo/video/document support
• mcp-kanban (8 tools): Programmatic CRUD for Kanban tasks
• mcp-vault (10 tools): Persistent document storage, SQLite FTS5 full-text search
• mcp-socialdata (5 tools): Twitter/X search, user profiles, engagement data

3. Agent Execution Layer

• Isolated claude CLI processes via node-pty
• Real-time output streaming per agent
• Status detected by parsing output patterns
• Autonomous execution supported (--dangerously-skip-permissions)
• Secondary project paths via --add-dir and git worktree support

Core Features

Parallel Agent Management

• Unlimited concurrent agents across projects
• Per-agent: isolated PTY, skill assignments, model selection (sonnet/opus/haiku)
• Lifecycle: idle → running → completed / error / waiting
• Persistent state across restarts (~/.dorothy/agents.json)

Super Agent (Orchestrator)

Meta-agent that programmatically creates, starts, stops other agents via MCP tools. Delegates by skills/capabilities, monitors progress, spawns temporary agents for one-off tasks. Remote-controllable via Telegram and Slack.

Automations

CI/CD-like AI workflows polling external sources on cron schedules:

Template variables ({{title}}, {{url}}, {{body}}, {{labels}}, etc.) inject item data into prompts. Content-hash deduplication prevents re-processing.

Kanban Task Management

Columns: Backlog → Planned → Ongoing → Done. Priority levels, 0-100% progress, manual/auto agent assignment, skill requirements. The kanban-automation service continuously matches tasks to agents by skill — creating agents if none match. JIRA automations auto-create Kanban backlog tasks, forming a closed loop from external issues to AI execution.

Scheduled Tasks

Cron-based recurring execution via platform-native schedulers (launchd on macOS, cron on Linux). Task definitions at ~/.claude/schedules.json, logs at ~/.claude/logs/.

Remote Control

• Telegram: Full fleet control via bot commands (/status, /agents, /start_agent, /stop_agent, /ask, /usage, /help). Messages without command → Super Agent. Multi-user.
• Slack: Equivalent via Socket Mode (no public URL needed). @mentions, DMs, thread-aware.

Vault

Persistent document storage: SQLite FTS5 full-text search, nested folders, cross-agent access, file attachments. MCP tools: create, read, update, delete, search, file attach, folder ops.

Google Workspace

Integration via gws CLI (Google Workspace CLI) as MCP server. Covers Gmail, Drive, Sheets, Calendar, Docs (R/W), Slides, Tasks, Chat, People, Forms, Keep. 100+ specialized skills available.

Skills & Plugins

One-click install from skills.sh and built-in marketplace. Categories: code intelligence (LSP for TS, Python, Rust, Go), external integrations (GitHub, GitLab, Jira, Figma, Slack, Vercel), dev workflows (commits, PR review).

Tech Stack

Configuration & Storage

All paths under ~/.dorothy/ and ~/.claude/: - App config: ~/.dorothy/app-settings.json - Persisted state: agents.json, kanban-tasks.json, automations.json, processed-items.json - SQLite: vault.db - Claude Code: ~/.claude/settings.json, ~/.claude/schedules.json, ~/.claude/logs/

Pricing

Free forever, open source, no account required.

Notable Design Decisions

1. MCP-first: All programmatic control exposed through five MCP servers — any MCP client can drive the system
2. node-pty isolation: Genuine PTY sessions per agent, not simple subprocess spawning
3. Content-hash deduplication: Automations track processed items by hash
4. Platform-native scheduling: launchd/cron rather than in-process timers
5. JIRA → Kanban → Agent closed loop: External issues auto-create Kanban tasks, auto-assigned to capable agents
6. Socket Mode for Slack: No public URL required
7. Persistent agent state: Agents survive app restarts