Back to Blog
Announcement

Introducing Hivemind: The Coordination Layer for AI Agents

When you run multiple AI coding agents on the same project, they step on each other. Hivemind fixes this with a shared event log and MCP server.

GB

Gabriel Bram

February 20, 20267 min read

The Problem

You open three terminals. Claude Code in one, Cursor in another, Codex CLI in the third. Each agent starts working on your project.

Within minutes:

1 Agent A and Agent B both start refactoring the auth module
2 Agent C decides to use Supabase. Agent A already chose Convex.
3 Agent B overwrites Agent A's changes to config.ts
4 You spend 30 minutes untangling the mess

This isn't a hypothetical. It's what happens every time you try to parallelize AI-assisted development. The agents are individually capable, but they have zero shared context.

What Hivemind Does

Hivemind is a persistent, append-only event log that gives every AI agent shared context. It's not an orchestrator, not a memory product, not a workflow engine. It's the coordination layer beneath all of them.

Claude Code
Agent 1
Cursor
Agent 2
Codex CLI
Agent 3
publish / query
Hivemind Event Log
Persistent · Append-only · Semantic Search · Advisory Locks

Every agent publishes structured events — decisions made, tasks completed, files locked, blockers found. Every other agent can query that log with semantic search to understand what's happening.

How It Works

The core primitive is simple: publish and query.

Agent 1 — publishes a decision
 hivemind_publish(
    channel: "backend",
    event_type: "decision.made",
    data: { description: "Switched to JWT auth" }
  )
 Published → evt_01JM8X...
Agent 2 — queries for context
 hivemind_query(query: "auth decisions")

Found 1 result: 0.94 similarity · decision.made · 2m ago "Switched to JWT auth"

Agent 2 doesn't need to know the exact words Agent 1 used. Semantic search matches by *meaning*, not keywords. "auth decisions" finds "Switched to JWT auth" at 0.94 similarity.

On top of publish/query, Hivemind provides:

  • Advisory file locks — prevent two agents editing the same file
  • Task state derivation — see what's active, completed, and blocked
  • Event triggers — react to patterns automatically
  • Knowledge base — persistent docs agents can reference
  • Approval workflows — human-in-the-loop for sensitive operations

Getting Started

Three commands:

Terminal
$ npm install -g @hivemindai/mcp-server
$ hivemind login
 Authenticated as you@email.com
$ hivemind init
 Project initialized

Add it to your Claude Code MCP config:

{ "hivemind": { "command": "hivemind-mcp" } }

That's it. Your agents now share context automatically. Hivemind is free during the beta.

launchmcpai-agents