Personal project · private network
Hive: a local multi-agent system
A hierarchical multi-agent system that runs on my own hardware. I talk to one master agent; it delegates to group leaders, they delegate to specialists, and the answers come back up the same path.
- Problem
- I wanted an agent team that runs on my own machine, that I can watch while it works, and that I can call from the tools I already use, instead of a black box billed per call.
- What I built
- A Go backend (standard library only) orchestrates the agents against a local LLM served by llama.cpp, streams every delegation over HTTP and SSE, and feeds a React and TypeScript dashboard. An MCP server over stdio exposes three tools (ask_master, list_agents, search_vault), so Claude Code or Copilot can call the system. Agent context comes from hybrid retrieval over an Obsidian vault: keyword search combined with embeddings generated by a local server. A job engine only marks a step as done when there is evidence: a written file, a saved note, a command that passed.
- Stack
- Go · React · TypeScript · HTTP / SSE · MCP (stdio) · llama.cpp · local embeddings · Tailscale
- Exists today
- Orchestration, live dashboard, MCP server, hybrid vault search, and the job engine, with Go and UI tests. It runs on a private Tailscale network, so there is no public demo.
- Limits and roadmap
- The local model is slow (agents share one GPU, so a round takes minutes) and can be wrong: answers are treated as drafts, and harder work can be handed to a cloud model. Agents take no external action without human approval.
