Build a web-based agent that performs real work, writing and running code or operating a real browser, with every action contained in a sandbox on Vultr. The system should plan tasks, dispatch them to isolated environments, and return verifiable output. Containment-first is required: an agent that only chats is a demo, one that executes safely is a product.
Tracks
Pattern A: Sandboxed code execution
- CSV Analyst: upload a spreadsheet, ask questions in plain English, get an executed chart plus the code.
- Prove-It Code Reviewer: run a snippet or PR and its tests in the sandbox, then report what happened.
- Chart Anything: natural language in, executed matplotlib or plotly image out.
- File Fixer: messy CSV or JSON in, cleaned file out, with an executed before/after diff.
- Self-Healing Runner: run code, read errors, patch and retry, with the retry loop visible on screen.
Pattern B: Sandboxed browser use
- Comparison Shopper: two or three product URLs in, screenshot-backed comparison table out.
- Form Filler: structured data in, a multi-step web form filled, with a screenshot trail and an approve-before-submit gate.
- Site QA Sweep: find broken flows and links in a web app, with screenshots and repro steps.
- Research with Receipts: a brief where every claim links to a screenshot of its source.
- Page Watcher: monitor listing or pricing pages, extract structured data, and diff against the previous day.
Containment focus areas
Address real safety and reliability challenges in agent execution:
- Process isolation: execution never touches the host or app runtime.
- Secret hygiene: no API keys or credentials inside the sandbox.
- Resource limits: time and memory caps on every run.
- Lifecycle discipline: reset or destroy the environment after each task.
Show how isolation makes autonomous execution shippable.
Required stack
- Vultr VM: the backend, acting as the central control and orchestration system.
- Vultr Serverless Inference: all agent LLM calls. OpenAI-compatible, base URL https://api.vultrinference.com/v1, vision-capable models available. Provisioning docs.
- Sandbox boundary: containers or throwaway instances on Vultr, never inside the app process. Docker strongly recommended; OpenSandbox, gVisor and E2B are good starting points.
- Vultr API (optional): spin up and destroy a throwaway instance per task.
Deliverables
- Public GitHub repository with setup and documentation
- Backend deployed on a Vultr VM, with LLM calls through Vultr Serverless Inference
- Public demo URL with a real product-style experience, not just a local demo
- Recorded demo video including one containment moment, such as the sandbox absorbing an rm -rf, an infinite loop, or a hostile web page
- Clear explanation of the architecture and use case
Who can join
Developers, AI/ML and platform engineers, and technical founders building agentic applications.
- Team size: 2 to 4 members. Solo participants are helped to join or form a team.
- What to bring: A laptop and a GitHub account. No special hardware needed.
- Skills: Working Python or TypeScript, API experience, and basic LLM familiarity.
Register on Luma first. If you are shortlisted, secure your spot on this page.