I built my custom agent manager. 7 months and 8,000+ Claude sessions later, it evolved into an application that replaced my Notion, Maps, and most of my daily apps
retained text
When I started using agents seriously, I also started to feel mentally exhausted when I was running more than 5-6 conversations in parallel. I felt this was due to agent manager UI making it hard to track the state of each conversation. So I started using Claude to build my custom agent manager. Rapidly I realized that the harness and primitives I developed were generic enough so they could be reused to build other applications. So I entered a quest of rebuilding personal versions of every application I was using, all inside a unified app. You can [try them in your browser](https://equin.ai/demo) (laptop): * Claude Desktop -> [https://equin.ai/demo/at/agents](https://equin.ai/demo/at/agents) * Notion/Obsidian -> [https://equin.ai/demo/at/pages](https://equin.ai/demo/at/pages) * Claude Design -> [https://equin.ai/demo/at/prototypes](https://equin.ai/demo/at/prototypes) * Maps -> [https://equin.ai/demo/at/places](https://equin.ai/demo/at/places) * Synthesia (a piano learning app) -> [https://equin.ai/demo/at/sonata](https://equin.ai/demo/at/sonata) * Finder -> [https://equin.ai/demo/at/files](https://equin.ai/demo/at/files) * My custom PaaS (runs only locally) * My custom window manager [https://equin.ai/demo/at/desktop](https://equin.ai/demo/at/desktop) Of course, all the original apps are prettier and have plenty of features that I don't have. But they are missing the specific features I wanted. I rarely ever use the original apps anymore. Creating the apps themselves was the fun/easy part. What was complicated was: * To design all the guardrails, automations, monitoring, iteration loops that allow me to fully trust Claude to update the app autonomously without breaking things. * All the performance optimisations, so this huge monolithic application will stay fast to build, fast to load, responsive,... no matter how big the app grows * Optimizing the process itself, so every iteration is faster than the previous one. Very often, the bottleneck was the UI more than the agents. I'm at the point where I can just throw most of my ideas in the system without switching context and everything gets updated in the background without even having to open my agent manager Now, I'm at the stage where the platform starts to be stable and mature enough that others could start to use it to build their own apps too. All code is fully open-source. It works on Mac and Linux, and you need a Claude subscription. My [end vision](https://equin.ai/about/vision) would be to create an open source platform of sharable building blocks/primitives/apps. So all users could shape their own custom self-evolving operating system and replace proprietary apps. I have plenty of ideas of things to improve. I still need to add mobile support, cross-device synchronization, and better sharing and collaborative capabilities. I've also learned a lot from this project about working with agents. If there's interest, I'll be happy to share more about my workflow and harness, so others can benefit from it. Website: [https://equin.ai](https://equin.ai) Repo: [https://github.com/equinai/equin](https://github.com/equinai/equin) Discord: [https://discord.gg/QYE8qSsfSQ](https://discord.gg/QYE8qSsfSQ)
document fetched 2026-11-19 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xzipn8/i_built_my_custom_agent_manager_7_months_and_8000