Hey Product Hunt 👋 I'm Sameer. I build developer tools and run multi-agent coding workflows daily — Claude Code, Codex, Cursor, sometimes all on the same project. Here's the problem I kept hitting: AI agents can write great code. But put three of them on the same codebase and chaos follows Agent A rewrites what Agent B just built. Nobody knows what's done. And I'm the one copy-pasting status between chat windows like a human message bus. I looked at every tool out there — TaskMaster, Mission Control, Vibe Kanban, Linear MCP, about 12 others. Each solved a piece. None solved the whole thing. And the MCP-based tools were burning 5,000-21,000 tokens per interaction just to check task status. That's insane So I built codepakt. It's a task board where agents interact via CLI instead of MCP: cpk task pickup --agent claude # claim next task (~250 tokens) cpk task done T-001 --notes "..." # mark complete, deps auto-resolve cpk docs search "auth" # query shared knowledge base ⚡ 250 tokens per interaction. Not 8,000. That's a 10-15x difference. What you get with one npm install 📦 🔄 Task board with dependencies, priorities, epics, and verification blocks 🔒 Atomic task pickup — two agents can't claim the same task (database transactions) 🔗 Auto-dependency resolution — when T-001 finishes, T-002 unlocks automatically 🧠 Shared knowledge base — agents write learnings, search when stuck 📊 Web dashboard at localhost:41920 — kanban board, agent sidebar, dark/light mode 📝 `cpk generate` creates AGENTS.md + CLAUDE.md so agents know the protocol on spawn 💾 Per-project SQLite — data lives with your code, like .git/ No Docker. No accounts. No API keys on the server. No LLM on the server. The server is dumb — it stores state and enforces concurrency. Agents do all the thinking 🧠 The moment I knew it worked ✅ I dogfooded it by building a snake game 🐍 with Claude and Codex working simultaneously on 13 tasks. Codex scaffolded the HTML/CSS/JS. Claude built the game logic. Codex came back for responsive CSS and QA. The QA task actually found a real bug 🐛 — keyboard handler checking reversal against the wrong variable. All coordinated through the shared board. I watched from the dashboard instead of being the bottleneck. How the approach evolved 🔄 The first design was cloud-first Postgres, a monorepo with three packages, server-side PRD decomposition via Claude API. All overengineered 😅 What shipped: single npm package, embedded SQLite, zero external dependencies. Running in 30 seconds ⏱️ I also stripped agent registration entirely. I originally built capability matching, agent rosters, the works. Then realized — the human already knows which agent does what. You say "Claude, pick up T-003." The server just needs to record who's working on what. So agents are just name strings now. Show up, claim work, report back. Simple 🎯 Where it's at 🚀 🔓 Open source (MIT): https://github.com/codepakt/cpk 📦 Install: npm i -g codepakt 🌐 Docs: https://codepakt.com It's early. v0.1. SQLite only. But it works — I'm using it on my own projects right now. Curious what coordination problems others are hitting with multi-agent workflows. What's still broken for you? 💬 — Sameer

codepakt

Task coordination for AI coding agents.

Task coordination for AI coding agents.