Proxy-lite: A lightweight, open-source, weighted autonomous assistant project

By /

Proxy Lite: If you want to use it after on-premises deployment, the endpoint address after the service runs is http://localhost:8008/v1. In addition, Proxy Lite defaults to the . Convergence AI is an open-source (mini) version of the agent/assistant framework. It allows the model to autonomously use tools (e.g., browser, click, search) to complete tasks in an environment.

1. Project introduction

  • Name: proxy-lite — “A mini, open-weights, version of our Proxy assistant.”
  • It is a lightweight, open-source version of Proxy Assistant (Weights Open).
  • The goal is to be an “autonomous agent”: the user sends a high-level task (such as “help me find a restaurant near a certain place and evaluate”), and the agent calls the tool, browser, analyzes the results, makes actions, and finally feeds back the results.

2. Key components & operation mechanisms

Here are the core components and workflow of proxy-lite:

Modules/ComponentsFunction / Function
Runner & RunnerConfigStart the agent’s main logic in Python, configure the environment, policies, models, tools, etc.
EnvironmentFor example, the “webbrowser” environment: the proxy can operate in the browser (click, type, jump, etc.).
ToolsActions that agents can use in the environment, such as BrowserTool,ReturnValueTool , etc.
Model / Client (client, model_id, api_base)Responsible for specific language reasoning, planning, tool calling, etc. You can call remote APIs or local services.
Message History (message_history)Structured conversations/observation history to give context to the model.
Tool Calls (tool_call)The model outputs calls in the form of instructions, such as “click”, “type”, etc., and the agent continues to interact after execution.

The operation process is roughly as follows:

  1. User gives a task (high-level intent)
  2. Runner wraps tasks in the message history
  3. The model is based on current observations + history + available tools to generate the next step “Think + Tool Call”
  4. Environment/Tool modules perform tool calls (e.g., browser clicks, searches, fetches page content)
  5. Feed back the observations, and then let the model decide the next step
  6. Repeat the above steps until the task completes or times out

In the model response, there will be <observation> … </observation>a structure like , <thinking> … </thinking><tool_call> … </tool_call>, to distinguish the three parts of “observation/reasoning/action”.

3. How to use / deploy

Here are some usage steps and deployment recommendations:

  • Run / install locally
    • Clone warehouse
    • Use make proxy(script provided by the project), or manually install it with pip / virtual environment
    • Install Playwright (for browser actions)
  • Invoke / use
    • Tasks like proxy "Find some markets near Kings Cross and tell me their ratings." this can be performed on the command line
    • You can also open a local web interface (web UI) for interactive presentations
    • Or in Python, use Runner Class + RunnerConfig Construct as a library call
  • Models / Interfaces
    • The default proxy-lite points to a presentation endpoint on a HuggingFace Spaces (not suitable for high-voltage use)
    • It is recommended that users deploy their own model endpoints, such as using vLLM to serveconvergence-ai/proxy-lite-3b  the model and enable the tool to call the parser --tool-call-parser hermes and other parameters
    • After deployment, specify your endpoint address on the client via or PROXY_LITE_API_BASE --api-base 

4. Advantages & Application Scenarios

Benefits / Value:

  • Autonomous execution capabilities: Given high-level goals, agents can call tools, browsers, searches, clicks, etc. to complete intermediate steps
  • Scalable/Modular: Environments, tools, and models can be replaced or expanded to adapt to different tasks
  • Open Source & Weighted Openness: This allows the community to experiment more freely than those large model systems that are completely closed-source
  • Lightweight / Relatively simplified: Although it is a simplified version, it may not have all the functions of the proxy, but it is enough to meet the prototype exploration of many proxy-based tasks

Application Scenarios:

  • Web automation tasks (e.g. “help me find a place” + automatic crawling, sorting, comparison)
  • Assistant / Agent Tasks (not just Q&A, but action + execution)
  • Research and prototyping: Want to test agent intelligence and tool call capabilities in an open environment
  • Assembled into more complex systems as submodules

5. Limitations/Precautions

The project also has some limitations and challenges:

  • Performance/Stability: The default demo endpoint is designed for demos and is not suitable for high loads or production use
  • Anti-bot / anti-bot challenge: Browser automation may be detected or blocked by the website / Captcha issue may occur
  • Security / Privacy: Automated proxies may access sensitive pages, submit forms, etc., requiring careful user control over authorization, credentials, etc
  • Task complexity limitations: Proxy-lite may struggle with extremely complex tasks that require a lot of conditional judgment and user interaction
  • Model capability bottleneck: As a lightweight version, it may not be as good as the more flagship version of the proxy system in terms of inference intelligence, tool call judgment capabilities, and understanding of complex contexts

Presentation endpoint for Hugging Face Spaces: https://convergence-ai-demo-api.hf.space/v1
GitHub ：https://github.com/convergence-ai/proxy-lite

Tubing:

Scroll to Top