I work at Archestra, and we're running a one-week online MCP Apps Hackathon: take one of your typical workflows and turn it into an app your team could use tomorrow.
- July 22–29, online, free, open worldwide, solo or teams
- $1,000 for the Most Useful App
- $750 for the Best Weird App (games and wonderfully weird interfaces have their own prize — usefulness not required)
- $500 + swag for Community Buzz: post your app anywhere public (LinkedIn, X, Reddit, Hacker News), tag or mention Archestra [@archestra_ai], most-liked post wins Register here
How building works
Archestra is an open-source AI platform. You describe the app in chat, it generates the interface, and the app talks to your real tools (GitHub, Jira, Slack, whatever your workflow lives in) through MCP (Model Context Protocol) servers, with your credentials. No deploy, no hosting, no login flow to build.
Setup is Docker plus an LLM API key — about 15 minutes. If you don't have a paid key - Cerebras has a free tier and local Ollama works. The full quickstart goes out to registrants on July 22.
What a submission looks like
A video of your app in action, the prompt you used, and which MCP servers it talks to. Every eligible submission goes into the public gallery on our website.
An app can be anything: a PR review queue, an invoice tracker, a contract review pipeline, a deploy dashboard, an LLM cost dashboard per project, etc. Or ignore all of that and build a game — last month someone on our team built a pub round ledger that our London office actually uses, which sits somewhere between the two prizes.
Register here — we'll send full brief, quickstart, and example apps to your inbox on July 22. Questions in the comments welcome, I'll be around.
Top comments (3)
Interesting approach. One thing I’m curious about: how do you handle MCP server trust boundaries and permission scoping? If an app can connect to GitHub, Jira, and Slack simultaneously, preventing unintended cross-tool data access seems just as important as generating the UI.
Thanks for the question! Some context first: Archestra is an open-source AI platform that runs your MCP servers in one place - it handles installing them, storing credentials, and routing every tool call through a gateway where it's logged. Apps are a layer on top: HTML interfaces that run in a sandboxed iframe and can only talk to the world through those MCP tools.
So to your question - an app doesn't get access to everything you've connected. Each app is assigned specific tools, and anything else is refused. Calls also run as the viewer, not the app's author: if a teammate opens your shared app without having connected Jira themselves, the call just fails with an auth prompt.
For exfiltration, the app's CSP is platform-pinned with connect-src 'none' — no fetch, no XHR. Assigned tools are the only way data moves, and those are logged at the gateway.
If you assign one app both a GitHub read tool and a Slack write tool, it can move data between them (usually that's the point of the app).
Here is our docs on apps: archestra.ai/docs/platform-apps
Thanks, Margaret! I appreciate you taking the time to explain it in detail and point me to the docs. The capability-based model and user-scoped execution answered most of my questions. Looking forward to experimenting with it during the hackathon.