A technician has created a "hive mind" system for Claude Code, based on a multi-agent architecture.

System Architecture

The system consists of seven specialized agents, each with a specific role (coder, tester, reviewer, architect, etc.). These agents are designed to coordinate on tasks, share persistent memory, and exchange context.

Key features include:

  • Seven agent types with different system prompts and capabilities.
  • SQLite + FTS5 for persistent memory, allowing agents to remember information between sessions.
  • A message bus for agent-to-agent communication.
  • A task queue with priority-based coordination.
  • Direct integration with Claude Code via MCP server.
  • Compatibility with Anthropic, OpenAI, or Ollama.

When a coder completes an implementation, the tester can query the shared memory to see what was built and write appropriate tests. The reviewer has access to the full context of the decisions made. The system uses TypeScript, better-sqlite3, MCP SDK, and Zod.

Challenges

Debugging seven agents communicating with each other can be complex. Sometimes, an agent may continuously assign tasks to itself, creating an infinite loop. The project is released under the MIT license and available on GitHub.