# Bandesh

- **Event:** [Meta Global AI Developer Hackathon](https://cerebralvalley.ai/e/meta-ai-developer-hackathon)
- **When:** Oct 26, 12:00 PM – Nov 4, 9:00 PM (EST)
- **Where:** Online
- **Team:** [Bandesh Degaon](https://cerebralvalley.ai/u/Nihal)
- **GitHub:** https://github.com/nihald2000/OpenHarness
- **Demo video:** https://youtu.be/zOXRqxofXsc
- **Gallery:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery
- **Page:** https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1596

Live Workspace: https://openharness-web-six.vercel.app

OpenHarness is an open-source, verifiable multimodal AI agent workspace built to close the reliability gap in agentic software and document engineering.

The Problem: Most AI coding agents write code and simply hope it works—declaring success without booting the server, clicking through the real UI, checking for visual layout collisions, or cleaning up background process failures. Running untrusted agent code or remote skills locally also exposes developers to severe security risks.

Our Solution: OpenHarness pairs Meta's 6-model multimodal stack with isolated Linux UID cloud sandboxes on Modal and a closed Plan-Build-Test-Fix state machine:
- Autonomous Live Browser & Human Takeover: Boots apps in <1s via a pre-allocated preview port pool (3010–3019), drives a live AI cursor in Agent View, streams frames over single-use HMAC WebSockets, and lets humans click "Take Control" anytime.
- Self-Healing Verification Gate: Runs repeatable 9-step Playwright end-to-end checks (`browser_check_def`), inspects background `process_logs`, catches latent port collisions (`EADDRINUSE`) and shutdown bugs (`SIGTERM`), patches code autonomously, and re-verifies before completion.
- Computer-Vision UI Proof & Multimodal Studio: Runs Meta SAM 3.1 (`visual_segment_audit`) on the live screen to flag bounding-box collisions (`layoutScore: 100`) and previews slide decks, PDFs, and formula-audited Excel workbooks inside Studio.
- Governed Extensibility & Rewind: Combines MCP server discovery (HTTP/SSE/stdio), a Skill Registry with approval gates for untrusted imports, and Shadow-Git checkpoints on every turn.

Built During the Hackathon (Oct 26 – Nov 4, 2026): All 6-model Meta API + Modal B200 vLLM open-weights integrations, repeatable `browser_check_def` self-healing loops, live WebSocket browser screencasting/takeover, SAM 3.1 visual QA, preview port pooling, MCP/Skill Registry approval gates, and Document Studio inspection were engineered during the Hackathon window.

## More from Meta Global AI Developer Hackathon

- [SamePage](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1593)
- [Ashura](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1594)
- [Mixray](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1595)
- [Touseef](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1597)
- [Janhavi Tupe](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1598)
- [Verify and Train on WhatsApp](https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1599)

---

Markdown version of https://cerebralvalley.ai/e/meta-ai-developer-hackathon/hackathon/gallery/1596. Site index for agents: https://cerebralvalley.ai/llms.txt · full text: https://cerebralvalley.ai/llms-full.txt