Open Source — Apache 2.0

One prompt.
A full AI engineering team.

Describe a feature in plain English. Hivemind deploys a PM, developers, reviewer, and QA — all working in parallel — and delivers tested, committed code.

$ npx create-hivemind@latest

If Claude Code is a developer, Hivemind is the engineering team.

🤖 Claude Code
🦞 OpenClaw
🧪 Codex
⌨️ Cursor
🐚 Bash
🌐 HTTP

If it can write code, it's hired.

How It Works

Three steps. One prompt to production code.

01

Describe the feature

"Add JWT authentication with a login page and protected routes."

Plain English. No config files. No YAML. Simple tasks execute instantly — complex ones get the full pipeline.

02

Watch the team work

Architect reviews the codebase. PM plans the DAG. Agents work in parallel. Send follow-up messages — they inject into the live DAG.

One living DAG per project. Tasks added or cancelled dynamically.

03

Get production code

Tested, reviewed (read-only critique), committed. Open your IDE and it's already there. Go lie on the couch.

Self-healing. If something fails, it fixes itself. If lint breaks tests, it reverts.

Features

Everything you need to run an AI engineering team.

🧩

LangGraph DAG Executor

Tasks execute in dependency order via a LangGraph StateGraph with SQLite checkpointing, self-healing retry, and dynamic task injection.

🔗

Dynamic DAG

Send new messages mid-execution — tasks are added or cancelled in the live DAG. Always one DAG per project, never parallel.

Adaptive Triage

Simple tasks skip the full PM + Architect pipeline and execute directly — reducing latency and token waste.

🏗️

Architect Agent

Pre-planning codebase review identifies patterns, conventions, and key files — giving the PM better context for planning.

🛡️

Read-Only Code Review

Reviewer critiques but never modifies code. Automated lint/format runs separately — reverted if it breaks tests.

🔄

Self-Healing

Failed tasks are classified by failure type and retried with targeted fixes. 5 stuck-detection signals with active escalation.

🔀

Artifact Flow

Agents pass typed artifacts (API contracts, schemas, test reports) to downstream agents as structured context.

🔒

Project Write Lock

Writer agents are serialized per project via asyncio.Lock. Reader agents run in parallel. No git conflicts.

🧠

Proactive Memory

Past failures and lessons are injected into agent prompts. The team learns across sessions.

💰

Zero Extra Cost

No API keys needed. Runs on your existing Claude Code CLI subscription. No token charges.

📱

Mobile Dashboard

Real-time streaming, DAG progress, file diffs, cost analytics — all from your phone.

🤝

Typed Contracts & Handoffs

Agents communicate via structured TaskInput/TaskOutput contracts and write detailed handoff documents for downstream agents.

Before & After

What changes when you add Hivemind to your workflow.

Without Hivemind With Hivemind
You ask Claude Code to build a feature. It works on one file at a time, loses context, and you babysit for hours. Describe the feature once. PM breaks it into a DAG, agents build in parallel, reviewer checks quality, code is committed.
For a full-stack feature, you manually coordinate backend, frontend, tests, review. Copy-pasting context between sessions. Artifact flow passes API contracts, schemas, and test reports between agents automatically.
An agent gets stuck in a loop. You kill it, lose context, start over. Self-healing detects stuck agents (5 distinct signals), reassigns, simplifies, or respawns — automatically.
You can't leave your desk. If you walk away, the agent stops or goes off track. Monitor from your phone. The dashboard streams everything in real-time. Walk away. Go to the couch.
Agents write buggy code and you only find out after merging. Read-only review gate catches issues before commit. If automated fixes break tests, they're reverted automatically.
Simple tasks go through the same heavy pipeline as complex ones, wasting tokens and time. Adaptive triage routes simple requests directly to execution, skipping PM + Architect overhead.
You send a follow-up message and it starts a whole new session, losing all progress. New messages inject tasks into the live DAG. One continuous execution, always growing — never parallel.

Meet the Team

Specialist agents, each built for a specific job.

🎯

Orchestrator

Central coordinator — triage, lifecycle management, DAG dispatch, dynamic task injection

🏛️

Architect

Pre-planning codebase review — identifies patterns, conventions, and tech stack for better planning

📋

PM Agent

Decomposes requests into a typed TaskGraph DAG with dependency wiring and role assignments

🧠

Memory Agent

Remembers decisions, patterns, and lessons from previous sessions to avoid repeating mistakes

🎨

Frontend Dev

Builds pixel-perfect UI components with React, TypeScript, and Tailwind CSS

⚙️

Backend Dev

Writes APIs, business logic, and server-side code with FastAPI and Python

🔧

Fullstack Dev

End-to-end implementation for simpler tasks routed through triage fast path

🗄️

Database Expert

Designs schemas, writes migrations, and optimizes queries for production workloads

🐳

DevOps

Configures Docker, CI/CD pipelines, and deployment infrastructure

🧪

Test Engineer

Writes tests in a strict TDD loop — runs them, fixes failures, proves they pass with real output

🔒

Security Auditor

Scans for vulnerabilities, enforces OWASP Top 10 compliance, and hardens your code

👁️

Reviewer

Read-only code critique — identifies issues without modifying code. Lint/format reverted if tests break.

Ready to ship features,
not babysit agents?

Open source. Self-hosted. No account required.

$ npx create-hivemind@latest