Drobek: bring your own AI agent, build an app, then take it outside

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 1: Touch Grass

What I Built

I built Drobek, an open-source platform where your own AI agent can create a small web app, give you a preview, and publish it at a URL you can share.

Yes, this is my project. The practical idea is simple: when someone asks an agent for a useful little app, they should be able to keep using it outside the conversation.

My take on Touch Grass is that anyone should be able to make a small website for something happening outside. Organising a walk, a picnic, an outdoor workshop, or an evening of stargazing? Ask your own AI agent for an event page with the meeting point, programme, what to bring, and a registration form. Drobek gives that page a home and a URL you can send to the people coming.

For example, you could start with: "Build a page for our Saturday stargazing meetup. Show the meeting place and programme, add a packing checklist and a signup form, and give me a preview before publishing." That is an example prompt, not a claim that this particular event has already happened. The forms module can store registrations and email the organiser; the app's design and content are yours to change.

The useful part is having a page that fits your event. Make it, share it, then go outside together.

There is already a concrete example running on it: Podzimní obloha, the autumn sky. It has Czech and English text, a Moon phase display, an autumn observing calendar, and a short guide to constellations. It even tells you to grab a blanket and tea and let your eyes adjust to the dark.

You can read what to look for, then put the phone away. That is the sort of app I want people to be able to make their own.

Demo

Open the autumn sky app to try an existing example. It demonstrates what people can publish with Drobek; it predates this challenge week.

The short video below introduces the platform and its agent-to-app workflow. It is an animated product explainer, rather than a recording of a timed build.

You can also browse the app gallery or connect your agent.

Code

The public Drobek repository contains the server, dashboard, MCP interface, SDK, platform modules, and self-hosting documentation. The core is AGPL-3.0. The agent plugins are MIT.

The repository has older history. The main rework I am putting forward here started on October 5; I have kept that history visible. The changelog documents the work, including version-history tools and streaming proxy support on October 5, followed by preview feedback, live data subscriptions, webhooks, and app analytics on October 6.

How I Built It

Drobek is the open-source tooling part of the AI workflow. The agent connects over MCP and calls tools such as create_app, write_files, and publish. The model runs in the agent's environment; Drobek does not bundle a model or require one particular provider.

When the agent writes files, the server compiles them with esbuild and returns compilation errors in the response. The agent can fix the code and try again. A successful build gets a preview URL. I can open it, test it, and decide when to publish that version.

The October work adds more to that loop. People can leave feedback on a preview, and the agent can retrieve and resolve it. It can inspect version history and keep a version for later. Apps can subscribe to data changes, and the proxy supports streaming from configured upstreams.

The app itself runs in the browser. Backend modules provide things like authentication, stored data, forms, files, and API access. Owners enter API secrets in the dashboard, and the proxy uses them server-side. The agent does not need the secret values.

There are deliberate limits: arbitrary app server code and per-app npm install are outside this model. A successful compile still needs a functional check. Preview and production also share the app's backend data, so rolling back code does not roll back records.

Why Does Open Innovation Matter?

The part between an AI-generated app and a usable URL is code you can inspect and run yourself. You can change the platform, add a backend module, or host it on your own server with Postgres, Redis, and Caddy.

MCP gives the agent a documented interface to that platform. The build loop and deployment do not depend on a hidden conversation inside one app builder. The choice of agent and model stays with the person building the app, including whether they use an open-source agent setup.

Self-hosting also gives you control over where the platform stores app data. The connected agent and its model provider have their own data handling and costs; choosing your own host does not make those disappear.

For this submission, the open contribution is Drobek itself: the infrastructure people can use to turn an idea into their own small app. Try the autumn sky example, and the source and self-hosting instructions are there if you want to build on it.