Hey Product Hunt 👋 I'm Santiago, a solo dev maintaining a multi-tenant cybersecurity platform. After three weeks away from the codebase, I'd open Claude Code and watch the agent re-explore the same files, make decisions I'd already made (sometimes contradicting them), and burn 40K+ tokens before writing a single useful line. The problem isn't speed. It's memory. So I built AgentNorth — an MCP server that: 🧠 Serves Claude curated context per module in a single call (instead of grep/read exploration) 📝 Logs every agent decision and change as an auditable trail 📊 Generates live documentation: dependency graphs, ERDs, module maps, decision timelines — written automatically as agents work 🔄 Bidirectional: agent reads context, leaves trace, next session starts informed Currently testing it on a real production repo (180K+ LOC). For the first time the docs are actually up to date — without anyone writing them. Open source, Apache 2.0. Works with any Claude plan and any MCP-compatible agent. 🔗 github.com/Santy1422/AgentNorth 🌐 agentnorth.io Would love feedback from anyone working with agents on large codebases. What's your biggest pain point with AI coding tools today?

AgentNorth — Open Source Pre-indexe MCP

Your agents share one source of truth

Your agents share one source of truth