a3

A3 Claude Plugin

A Claude Code plugin for running Agent-as-Annotators (A3) web agents in BrowserGym's interactive chat mode.

Installation

/plugin marketplace add xhluca/a3-claude-plugin
/plugin install a3

To update to the latest version:

/plugin marketplace update
/plugin update a3

Usage

/a3:chat

With options:

/a3:chat --vllm-url https://your-vllm-endpoint.com/v1 --model your-model-name --start-url https://www.google.com

What it does

1. Clones and installs the agent-as-annotators package with uv
2. Auto-detects the model served at the vLLM endpoint
3. Sets up a virtual display if running on a headless server (no $DISPLAY)
4. Launches BrowserGym's interactive chat mode with the A3 agent

Defaults

Option	Default
--vllm-url	https://a3-qwen-vllm.mcgill-nlp.org/v1
--model	Auto-detected from endpoint
--start-url	https://www.google.com

Links

• A3 Paper
• A3 GitHub
• A3-Synth Dataset
• A3 Models

Running locally on Mac

You can self-host the A3 model on your Mac using llama.cpp with a 4-bit GGUF quantization from mradermacher/A3-Qwen3.5-9B-GGUF.

1. Install llama.cpp

brew install llama.cpp

2. Download the model

Download the Q4_K_M quantization (~5.7 GB) and the multimodal projector:

huggingface-cli download mradermacher/A3-Qwen3.5-9B-GGUF A3-Qwen3.5-9B.Q4_K_M.gguf --local-dir .
huggingface-cli download mradermacher/A3-Qwen3.5-9B-GGUF mmproj-A3-Qwen3.5-9B-f16.gguf --local-dir .

3. Start the server

llama-server \
  --model A3-Qwen3.5-9B.Q4_K_M.gguf \
  --mmproj mmproj-A3-Qwen3.5-9B-f16.gguf \
  --port 8080

The server will start at http://localhost:8080 with an OpenAI-compatible API.

4. Use with the plugin

/a3:chat --base-url http://localhost:8080/v1

Example

/a3:chat --base-url http://localhost:8080/v1 --start-url https://flights.google.com

Then in the chat, type:

Show me the 3 cheapest flights from Montreal to Rio de Janeiro for April 22 to 28, under 2 layovers

License

MIT