WebVoyager SOP

Appearance

Test ​

Status: NEEDS-ACCOUNT (GitHub clone + API key required) ​

Classification: NEEDS-ACCOUNT

WebVoyager is not available as a published pip package.

  • pip show webvoyager returns: not found

WebVoyager is a GitHub-only Python project at github.com/MinorJerry/WebVoyager. It requires a manual clone, pip install -r requirements.txt inside a conda environment, Playwright browser setup, and a valid OpenAI API key (GPT-4V access required) before any run is possible.

What is required ​

  1. Clone the repository:

    bash
    git clone https://github.com/MinorJerry/WebVoyager.git
cd WebVoyager

  2. Create and activate the conda environment:

    bash
    conda create -n webvoyager python=3.10
conda activate webvoyager
pip install -r requirements.txt

  3. Install Playwright browsers:

    bash
    playwright install chromium

  4. Obtain an OpenAI API key with GPT-4V access.

Smoke test (after manual install) ​

WebVoyager has no --version flag and no standalone unit-test command. The minimal end-to-end smoke test is to run the agent against the bundled test task file with a single iteration cap.

bash
conda activate webvoyager

python run.py \
    --test_file ./data/tasks_test.jsonl \
    --api_key YOUR_OPENAI_API_KEY \
    --headless \
    --max_iter 1 \
    --max_attached_imgs 1 \
    --temperature 0 \
    --seed 42

Setting --max_iter 1 stops after one browser interaction per task, which is enough to confirm Chrome launches, the model responds, and the results directory is created without running a full expensive benchmark.

Expected success output ​

When the agent initializes successfully you will see log lines similar to:

Start processing 25 tasks...
Task 1: https://www.google.com | "Search for ..."
Observation: [screenshot taken, page loaded]
Action: CLICK [element_id]
...
Result saved to results/examples/task_1/

The results/examples/ directory will be populated with subdirectories, each containing:

  • interact_messages.json - full model conversation log
  • One or more .png screenshots captured during the run

Verifying the environment only (no API key needed) ​

To confirm Python dependencies installed correctly without spending API credits, run:

bash
conda activate webvoyager
python -c "import selenium; import openai; from PIL import Image; print('deps OK')"

Expected output:

deps OK

If any import fails, re-run pip install -r requirements.txt inside the activated conda environment.

Notes ​

  • A real OpenAI API key with GPT-4V access is required for a full end-to-end run. Expect costs of roughly $0.01 to $0.05 per task depending on screenshot count and task complexity.
  • The smoke test is not runnable in a sandboxed CI environment because it requires a live Chrome install and outbound network access to OpenAI and the target websites.
  • tested=false: there is no pip package and no one-liner install. Commands are documented from the repository source and were not executed in the build environment.