Hearth lets families build custom apps inside a shared workspace

Imagine a shared Obsidian vault for your entire household, but with an LLM agent that actually understands the context of your family's schedules, notes, and rituals. That is essentially what Hearth does. Instead of just storing static information, the agent in Hearth can spin up functional apps based on the data you have already logged. For example, if you have travel notes and calendars scattered around, the agent can build a dedicated travel app that runs right inside the workspace to manage the chaos.

What makes this technically interesting is that it is not just a wrapper around a chat interface. It is built on a library called Playground, which is designed specifically for collaborative AI coding harnesses. The architecture relies on a few core primitives:

• Synchronized files: ensuring everyone in the household sees the same state in real time.
• App code sandboxing: this is the critical part that allows the agent to generate and execute code for those custom apps without crashing the whole environment.
• Policy layer: managing who can do what within the shared space.
• Agent integration: an LLM that has read and write access to the shared context to facilitate the app-building process.

The versatility of this shared workspace plus agent model is clear when you see it applied elsewhere. While Hearth is aimed at families, the same underlying framework is being used to build Bear for construction project management. It proves that the AI workflow of contextual notes to agent-generated tool to integrated app works across different domains, whether you are coordinating a vacation or a building site.

Since this is still in beta, the focus has been on isolation and least-privilege access. However, for anyone wanting a real-world deployment of a personal LLM agent, it is a great example of how to move beyond simple prompting and into actual utility. If you are into prompt engineering or building your own AI workflow, the idea of an agent that creates its own UI to solve a specific data problem is definitely the right direction.

For those interested in the deeper technical architecture or how the assistant is structured, there is a detailed breakdown here:

https://bear.build/posts/hearth-build-your-own-assistant/

The project is heading toward being open source soon, which will be the real test for those of us who are hesitant to put sensitive family data into a closed beta. Until then, it serves as a solid proof of concept for how a shared knowledge base can evolve into a suite of custom, agent-built tools.