Dorothy is a free and open-source desktop application for orchestrating Claude Code, Codex, Gemini, Grok, and local AI agents. It is designed to address the limitation that AI CLI tools can only run one agent at a time in one terminal, allowing users to run more than 10 agents simultaneously, deploying, monitoring, and debugging across projects and codebases.
Purpose Dorothy is aimed at developers and teams that need to manage multiple AI agents in parallel. Through a unified graphical interface, it organizes scattered CLI agents into a schedulable, monitorable, and automatable resource pool.
Key capabilities Parallel agent management: Each agent runs in an independent PTY terminal session, supports running an unlimited number of agents simultaneously, can be assigned skills, allows selection of models (sonnet, opus, haiku) and project context, sends interactive input in real time, and streams output. It supports agent lifecycle management (idle → running → completed / error / waiting), adding multi-repository context via --add-dir, Git worktree branch-isolated development, preserving agent state after application restart, and unattended autonomous execution mode (--dangerously-skip-permissions). Super Agent (orchestrator): A meta-agent that programmatically creates, starts, and stops other agents through MCP tools, delegates tasks based on capabilities and skills, monitors progress, captures output, and handles errors. It can also respond to Telegram and Slack messages for remote orchestration, and temporarily create and clean up agents for one-off tasks. Skills and plugin system: Extend agent capabilities through skills.sh and the built-in plugin marketplace, including LSP plugins for TypeScript, Python, Rust, Go, and more, external integrations such as GitHub, GitLab, Jira, Figma, Slack, and Vercel, as well as development workflows such as commit commands and PR reviews. Supports installing skills individually for each agent. Automation pipelines: Poll external sources (such as GitHub PRs and JIRA issues), detect new or updated items, deduplicate by content hash, generate a temporary agent for each item, inject template variables, execute autonomously, and deliver results to Telegram, Slack, or GitHub comments. JIRA automation also automatically creates kanban tasks in the to-do list. KanbanTask management: Tasks flow between backlog → planned → ongoing → done columns, supporting priority (low/medium/high), progress tracking (0-100%), agent assignment, tag filtering, and skill requirement matching. The kanban-automation service automatically matches and assigns tasks to available agents. Scheduled tasks: Autonomously run Claude Code based on cron scheduling, using launchd on macOS and cron on Linux. Task definitions are stored in ~/.claude/schedules.json, scripts are generated in ~/.dorothy/scripts/, and logs are located in ~/.claude/logs/. Remote control: Launch agents, view status, and delegate tasks via a Telegram bot or Slack (supporting @mentions, direct messages, Socket Mode, and thread awareness). Telegram also supports sending photos, videos, and documents via mcp-telegram. Persistent document vault (Vault): A Markdown document storage system that agents can read from and write to across sessions, supporting titles, content, tags, and file attachments, with support for nested folders and SQLite FTS5-based full-text search, and accessible across agents. External data and integrations: Search tweets and retrieve user profiles and engagement data via the SocialData API; access Gmail, Drive, Sheets, Docs, Calendar, etc. via the Google Workspace CLI (gws); five built-in MCP servers (mcp-orchestrator, mcp-telegram, mcp-kanban, mcp-vault, mcp-socialdata) provide more than 40 tools.
How to use Users can download the latest version from GitHub Releases. Claude Code CLI must be installed first (npm install -g @anthropic-ai/claude-code), and GitHub automation also requires the GitHub CLI (gh). If macOS reports that the app is damaged, you can run xattr -cr /Applications/Dorothy.app. You can also build from source: on macOS, output release/mac-arm64/Dorothy.app (Apple Silicon) or release/mac/Dorothy.app (Intel), including a DMG installer; in development mode you can open http://localhost:3000，但智能体管理与终端功能需 Electron app. For Telegram, you need to create a bot via @BotFather and paste the token; for Slack, you need to create an app at api.slack.com/apps and configure Socket Mode and the corresponding permissions. For Google Workspace, you need to install the gcloud CLI and gws CLI, and complete authorization in the settings.
Use cases Suitable for developers who need to manage multiple AI coding agents at the same time, teams that want to integrate AI agents into CI/CD-style automation workflows, and users who need to remotely control agents from mobile devices, build shared knowledge bases, or perform social media data analysis.
Limitations The project is open source and uses the MIT license. Agent management features require the Electron desktop app and are unavailable in browser development mode. Some integrations depend on external services (for example, the SocialData API requires obtaining your own key, and Google Workspace requires completing OAuth setup). Other details related to cost, platform, or effectiveness are not currently provided.
