A mini, open-weights version of the Proxy autonomous assistant, this project enables users to build and deploy AI agents capable of performing web-based tasks. It is designed for developers and researchers looking to integrate autonomous web navigation and task completion into their applications.

How It Works

Proxy Lite employs a modular architecture, allowing for interchangeable environments, solvers, and agents. The core functionality relies on a language model that generates "Observe, Think, Act" sequences. It uses a web browser environment powered by Playwright, which interacts with web pages by identifying elements via mark_id. Context-window management is handled by discarding older observations to reduce token count, while still retaining historical context through model reflections.

Quick Start & Requirements

• Install via make proxy or manually with pip install uv, uv venv, uv sync, and uv pip install -e ..
• Requires Python 3.11.
• Playwright needs to be installed via playwright install.
• For local serving with vLLM, install vLLM and transformers with uv sync --all-extras. Qwen-2.5-VL requires installation from source.
• Official docs: https://github.com/convergence-ai/proxy-lite

Highlighted Details

• CLI usage: proxy "Task description."
• Local web UI: make app
• Python API: Runner class for programmatic task execution.
• Supports vision input for multimodal task understanding.
• vLLM serving command provided for local deployment.

Maintenance & Community

• Developed by Convergence AI.
• Citation available in BibTeX format.

Licensing & Compatibility

• License: Not explicitly stated in the README.

Limitations & Caveats

The project is not designed for direct user interaction or tasks requiring credentials; it excels at autonomous task completion. Anti-bot measures may block execution, and network proxies are recommended. The demo endpoint is not suitable for production use.