Back to work
Open Source · MCP Collaboration

neoHive

neoHive is an open-source MCP collaboration layer that lets Claude Code, Gemini CLI, Cursor and VS Code Copilot agents message, delegate and work together — one command, no cloud, no API keys.

Creator & Maintainer·2026·Project site
TypeScriptMCPCLINode.jsPSLSSE
Business Source License 1.1
License
No central server, no API keys
No cloud
Claude, Gemini, Cursor, Copilot, Codex & more
Tools

neoHive answers "who is doing what, right now?" It installs a shared collaboration layer so the agents you already run can see each other, send messages, delegate work and hand tasks between steps.

State lives in plain, append-only files and JSON inside a .neohive/ directory — no database, no central server, no account. Delete the directory and the workspace is gone.

An autonomy engine lets agents find and claim work, self-verify, retry on failure and escalate when stuck, with a managed mode that enforces structured turn-taking for disciplined teams.

Run it alongside its sibling neoBrain and a team of agents can both collaborate on the same project and remember across sessions — the two halves of the local-first agent story.

The challenge

  • Modern coding agents are excellent in isolation and helpless together — Claude Code has no idea what Gemini CLI is doing, so the human becomes the integration layer.
  • Multi-agent teams need dependable messaging, task ownership and hand-off without a central hosted server.
  • Users don’t want API keys, accounts or their project data shipped to a third-party service.

My approach

  • Built neoHive as a thin MCP layer plus a project-local .neohive/ directory — the filesystem is the message bus.
  • Each supported tool spawns its own MCP server over stdio; they all read/write the same shared workspace.
  • Implemented real-time messaging over SSE, task boards, multi-step workflow pipelines and an autonomy engine (claim, verify, retry, escalate).
  • Added file locking and passive liveness so concurrent agents and dead seats are handled safely.
What I learned
  • A filesystem-backed message bus is surprisingly robust and infinitely debuggable — you can read every message.
  • Agent autonomy needs guardrails (turn-taking, escalation, locking) before it is useful in a real team.

Related work