---
title: OpenVLHarness
emoji: 🔍
colorFrom: indigo
colorTo: blue
sdk: static
pinned: false
short_description: Visual tool-use agent on your own model
datasets:
  - uclanlp/OpenVLHarness-Evaluation-Datasets
---

# OpenVLHarness

**Paper:** [OpenVLHarness: Real World Visual Agentic Tool-Use with Multimodal Memory](https://arxiv.org/abs/2610.26860)

Interactive demo of [OpenVLHarness](https://openvlharness.github.io/): ask questions about images with your own model (OpenAI, or an OpenAI-compatible endpoint with tool calling).

The agent loop and perception tools run on our lab server.

This page is generated by `scripts/build_space.py` in the project page repository.

## Shared backend

This Space and [the original demo](https://huggingface.co/spaces/bryanzhou008/OpenVLHarness-Demo)
use the same lab-hosted agent and perception services. This repository hosts the static frontend only.
The frontend discovers the current backend URL from the original demo page on each load, so updating
that page when the tunnel changes updates both demos. No second GPU deployment is started.

Initial frontend copied from `bryanzhou008/OpenVLHarness-Demo` at
`6effcacfeae839962d8912a01eae4d918ac28ef8`. For future frontend updates, copy the public assets
from that Space while preserving this Space's name and the shared backend-discovery URL.