Blame for README.md

---
license: apache-2.0
pretty_name: WebsiteBench
language:
- en
tags:
- benchmark
- agents
- web-agents
- website-reconstruction
- harbor
viewer: false
---

# WebsiteBench

Given a production-level website they can explore through a browser, agents must
reverse-engineer a full-stack version that matches its appearance and behavior as
closely as possible.

This dataset holds the 70 WebsiteBench tasks in the native task format of the
[Harbor framework](https://github.com/harbor-framework/harbor) (version 0.25.0).
The same task contents are published on [Harbor Hub](https://hub.harborframework.com/)
as `websitebench/WebsiteBench@v1`; every task has the same Harbor content hash in both
places. Code, documentation, and the evaluation commands live on
[GitHub](https://github.com/TIGER-AI-Lab/WebsiteBench).

- Project page: [website-bench.com](https://website-bench.com/)
- Code: [TIGER-AI-Lab/WebsiteBench](https://github.com/TIGER-AI-Lab/WebsiteBench)

## Download and run

No login is needed. Install the `websitebench` package, which pins Harbor 0.25.0, and
download with it. It checks every task against its pinned Harbor content hash and restores
the executable permissions that Hugging Face does not store:

```sh
pip install websitebench
websitebench download --source hf --task medium
harbor run --path ./local-tasks/medium --agent oracle
```

Or from a clone of the GitHub repository:

```sh
git clone https://github.com/TIGER-AI-Lab/WebsiteBench.git
cd WebsiteBench
uv sync
uv run python scripts/download_tasks.py --source hf --task medium
```

Omit `--task` to download all 70 tasks (9.7 GB) into `local-tasks/`. Hugging Face
rate-limits anonymous downloads, so a full download can pause for a few minutes at a time;
logging in with `hf auth login`, or setting `HF_TOKEN`, raises the limit.

Every task requests 8 CPUs, 8 GiB of memory and 20 GiB of disk for each agent and verifier environment; see the GitHub README for
requirements. The GitHub
[documentation](https://github.com/TIGER-AI-Lab/WebsiteBench/tree/main/docs)
describes coding-agent runs, batch runs, and results.

## Contents

| Path | Contents |
| --- | --- |
| `tasks/<task>/instruction.md` | The reconstruction task given to the agent |
| `tasks/<task>/task.toml` | Harbor configuration, resources, and network policies |
| `tasks/<task>/environment/` | Agent environment, browser gateway, mailbox, and the reference website |
| `tasks/<task>/solution/` | The bundled oracle implementation installed by `solve.sh` |
| `tasks/<task>/tests/` | The separate verifier, its oracle runtime, and the hidden test suite |
| `manifest.json` | SHA-256, size, and mode of every file, and each task's Harbor content hash |

In four tasks with very large static sites (`backcountry`, `columbia`, `eatthismuch`,
`warbyparker`), each application copy stores its `clone/static/` and `clone/backend/content/`
files in one `websitebench-static.tar`. The task's Dockerfiles, `tests/test.sh` and
`solution/solve.sh` unpack it, so the files inside the containers are the same as in an
unpacked copy.

## Scoring

Each task has 302 to 2,016 hidden test cases, 46,916 in total. A case passes when its
functional checks pass and every visual checkpoint's flat-baseline-normalized RGB SSIM
reaches 0.99. A task's reward is the number of passed cases divided by the number of cases.

## License and terms

The benchmark code, task definitions, and verifier are released under the
Apache-2.0 license. The reference websites reproduce third-party sites for
evaluation; their content remains the property of its owners and is provided for
non-commercial research only.

The task definitions carry a `harbor-canary` GUID. Please exclude documents containing
it from training data.

## Citation

```bibtex
@misc{zhang2026websitebench,
  title  = {WebsiteBench: Can AI Agents Rebuild Websites through Browser Exploration?},
  author = {{WebsiteBench Authors}},
  year   = {2026},
  url    = {https://github.com/TIGER-AI-Lab/WebsiteBench},
}
```