An open-source multi-agent framework developed by Mariano Pardo (GitHub: marian2js). True to its concept of “running your AI team like a company,” it arranges agents with roles such as CEO, manager, engineer, and designer into an org chart, where higher-level agents break down work and delegate it to their reports. The agent runtime is not fixed: the agents that do the actual work can run on Claude Code, Codex, Cursor, GitHub Copilot CLI, and others. Released under the MIT license, its biggest selling point is that it can be fully self-hosted on your own machine or server.
Key Features
- Agents organized as an org chart: Each agent is defined with a SOUL (personality), ROLE, VISION, MISSION, and VALUES, and is linked to a manager and reportees. You build an AI team using almost the same structure as a human company’s org chart
- Hierarchical task delegation: A coordinating agent called “The Goat” receives a project, breaks it into tasks, and assigns them along the organizational structure. Managers can subdivide further and hand work to their reports, enabling multi-level delegation
- Mixed runtimes: The manager layer runs on OpenClaw, while specialists handling implementation can be assigned to Claude Code, Codex, Cursor, OpenCode, GitHub Copilot CLI, Lovable, and more. You can pick the right tool for each area of strength
- CLI-first plus a dashboard: Start with three commands ─
npm i -g openclaw opengoat,openclaw onboard, andopengoat start─ then review the org chart, sessions, tasks, and activity logs on the dashboard athttp://127.0.0.1:19123 - Full self-hosting and ACP support: A Docker image is provided, and installation from source is supported. It also supports ACP (Agent Client Protocol), so external agent clients can connect to it
- Session and task continuity: Conversation sessions persist across multiple turns, and the platform includes task management and an extensible skill system
Pricing
| Plan | Price | What’s included |
|---|---|---|
| Open source (MIT license) | $0 | All features. No subscription and no hosted lock-in |
Pricing is as of August 2026. OpenGoat itself is free, but you still pay separately for the agent runtimes you actually run (Claude Code, Codex, Cursor, and so on) and for the model API usage behind them. Check the official site for the latest information.
Pros & Cons
✅ Pros
- No subscription and no vendor lock-in, so you can bring multi-agent operations into your own environment as-is
- Because you can write out “who instructs whom” as an org chart, complex agent coordination is expressed as structure rather than code
- You are not tied to a single execution agent. Claude Code or Codex subscriptions you already hold can be folded in as working members
- The dashboard lets you follow task progress and agent activity logs, so work does not become a black box
- The MIT license leaves plenty of freedom to embed it in internal tools or modify it
⚠️ Cons
- It requires Node.js 20.11 or later, CLI operation, and authentication setup for each runtime, which is a high barrier for non-engineers
- Since it assumes self-hosting, server management, updates, and incident response are your responsibility
- Running several agents in parallel tends to increase model API consumption, so cost management is needed
- It is a relatively new project, and specifications or CLI commands may change between releases. Checking the CHANGELOG is essential for production use
- Japanese-language articles and case studies are still scarce, so you will largely rely on the official documentation
Comparison with Similar Services
| Item | OpenGoat | CrewAI | MetaGPT | Claude Code subagents |
|---|---|---|---|---|
| Core idea | A company org chart (managers and reports) | Crews of role-based agents | Software company workflow (PM, design, implementation) | A parent agent calls purpose-specific children |
| Main interface | CLI plus dashboard | Python code | Python code | Terminal (config files) |
| Execution runtime | Choose Claude Code / Codex / Cursor, etc. | Calls LLM APIs directly | Calls LLM APIs directly | Fixed to Claude |
| Self-hosting | Yes (Docker supported) | Yes | Yes | Runs locally |
| License and cost | MIT, free | Open-source edition available (commercial edition separate) | Open source | Included in your Claude plan |
Who Is It For
- Developers who have outgrown a single coding agent and want to split work across several agents
- Teams that want to keep multi-agent workloads entirely on their own servers instead of handing data to an external SaaS
- People already using Claude Code or Codex who are looking for an upper layer to coordinate them as “reports”
- Anyone who wants to design agent role separation and delegation flows as an organizational structure rather than as code
- Those who want to read through an open-source agent platform and adapt it to their own operations
Summary
OpenGoat is a framework built on the idea of treating AI agents not as one all-purpose assistant but as a group of employees with defined roles. Because it borrows the familiar metaphors of an org chart and delegation, coordinating multiple agents becomes intuitive to design. On the other hand, it is a self-hosted CLI tool, so environment setup and operational management fall on the user. A realistic starting point is to launch a small organization locally ─ one coordinator plus one or two implementers ─ and connect the coding agents you already use.