5dive: Run AI Agent Companies on Your Own Server

What changed

The 5dive project provides a framework for orchestrating multiple AI agents on a user-owned server, operating as a self-contained company. Instead of relying on external platforms or complex protocols, 5dive utilizes fundamental Linux utilities such as systemd for agent supervision and bash for command-line interaction. Each AI agent runs as a distinct Linux user, ensuring isolation. These agents can be named and configured to use various official coding CLIs, including models from Anthropic (Claude), OpenAI (Codex), and others. The system supports a shared SQLite task queue for managing work and facilitates automated handoffs between agents. Human intervention is designed to be minimal, primarily for critical decisions, with notifications sent to a user's phone. The project emphasizes running on modest hardware, with a single-agent setup capable of operating on a 1GB VM. The latest release, v0.17.1, continues this approach. The installation process is streamlined, typically involving a curl | sudo bash command followed by initialization, with options for non-interactive setup.

Supported Agent Types and Configuration

5dive supports a wide array of agent types, each potentially connecting to different model families and authentication methods:

• claude: Supports Anthropic Claude or compatible endpoints, using OAuth, API keys, or --provider flags. Channels include Telegram and Discord.
• codex: For OpenAI Codex, using OAuth or API keys, with Telegram channel support.
• antigravity: Integrates with Google Antigravity via Google OAuth, supporting Telegram.
• grok: Connects to xAI Grok using OAuth or API keys, with Telegram support.
• devin: For Cognition Devin, requiring Devin account OAuth.
• hermes, openclaw: Community-built harnesses supporting multiple providers like OpenRouter, Anthropic, Google, and others via API keys, with Telegram and Discord channels.
• opencode: Supports OpenCode via API keys, with Telegram channels.
• pi: A third-party harness using API keys or --provider flags, supporting Telegram.

Configuration allows for custom endpoints using --base-url for Anthropic-compatible services, enabling integration with self-hosted models or less common vendors. Users can specify models, manage API keys through profiles, and rebind agents to different accounts or providers. The project also supports plugins, such as the 5dive-council for structured decision-making.

Why it matters for builders

This project empowers builders with a robust, self-hostable solution for managing complex AI agent ecosystems. It provides a foundational layer for creating autonomous workflows without relying on proprietary platforms, offering enhanced control over data privacy and operational costs. The use of standard Linux tools simplifies deployment and maintenance, making advanced agent orchestration accessible even on resource-constrained environments.

Practical impact

Builders can leverage 5dive to deploy and manage their own AI agent companies, starting with simple setups and scaling to complex organizational structures. The project's emphasis on using official CLIs with user-provided keys means direct access to AI models without intermediary services. Developers can experiment with creating custom agent roles, defining inter-agent communication patterns, and integrating with their existing infrastructure through the command-line interface. The ability to run agents as systemd services ensures persistence and reliability. The project's MIT license encourages broad adoption and modification.

Caveats and source limits

The source material describes the architecture and capabilities of 5dive but does not provide specific performance benchmarks or independent validation of its claims regarding agent efficiency or resource usage. While it mentions support for various agent CLIs, the exact compatibility and performance with all third-party harnesses and custom endpoints may require further testing by users. The project is described as a "fresh release" with a recent version number (v0.17.1), indicating it is still under active development and may evolve significantly. Pricing for the managed VM option is not detailed in the provided excerpts, and users are responsible for their own agent CLI subscriptions or API key costs.