pierretokns/mcp-browser-automation-v2

MCP Browser Automation Training Dataset v2

Training data for teaching Vision-Language Models (VLMs) to use Claude Chrome MCP tools for browser automation.

Version History

Version	Date	Changes
v2	2026-01-28	Fixed tool names to use proper MCP namespaced format (mcp__claude-in-chrome__*)
v1	2026-01-27	Initial release with generic tool names

What's New in v2

Critical Fix: All tool names now use the correct MCP namespaced format that Claude Code expects:

v1 (Wrong)	v2 (Correct)
navigate	mcp__claude-in-chrome__navigate
computer	mcp__claude-in-chrome__computer
screenshot	mcp__claude-in-chrome__screenshot
etc.	etc.

This ensures models trained on this data will generate tool calls that Claude Code can actually execute.

Dataset Description

This dataset contains 1,980 examples of browser automation tasks with corresponding MCP tool calls. Each example includes:

• A system prompt describing available tools
• A user request (natural language task)
• Assistant responses with <tool_call> blocks
• Simulated tool responses
• Task completion messages

Tools Covered (16 total)

• mcp__claude-in-chrome__navigate: Go to URLs
• mcp__claude-in-chrome__read_page: Get accessibility tree
• mcp__claude-in-chrome__find: Find elements by text/selector/role
• mcp__claude-in-chrome__computer: Click, type, scroll, keyboard
• mcp__claude-in-chrome__form_input: Fill form fields
• mcp__claude-in-chrome__get_page_text: Extract text content
• mcp__claude-in-chrome__screenshot: Capture screenshots
• mcp__claude-in-chrome__javascript_tool: Run JavaScript
• mcp__claude-in-chrome__tabs_context_mcp: Get tab info
• mcp__claude-in-chrome__tabs_create_mcp: Create new tab
• mcp__claude-in-chrome__gif_creator: Record interactions
• mcp__claude-in-chrome__upload_image: Upload files
• mcp__claude-in-chrome__read_console_messages: Read console
• mcp__claude-in-chrome__read_network_requests: Read network
• mcp__claude-in-chrome__shortcuts_list: List shortcuts
• mcp__claude-in-chrome__shortcuts_execute: Run shortcuts

Formats

Two formats are provided for different training frameworks:

Text Format (mcp_training_text/)

Qwen-style chat template with <|im_start|> tokens. Suitable for:

• MLX-VLM LoRA training
• Text-only models
• Models using Qwen chat template

{
  "text": "<|im_start|>system\n...<|im_end|>\n<|im_start|>user\nGo to google.com<|im_end|>\n<|im_start|>assistant\n<tool_call>\n{\"name\": \"mcp__claude-in-chrome__navigate\", \"arguments\": \"{\\\"url\\\": \\\"https://google.com\\\"}\"}\n</tool_call><|im_end|>..."
}

Messages Format (mcp_training_split/)

OpenAI function calling format. Suitable for:

• OpenAI-compatible training pipelines
• Models using function calling format

{
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "Go to google.com"},
    {"role": "assistant", "content": null, "tool_calls": [
      {"id": "call_123", "type": "function", "function": {"name": "mcp__claude-in-chrome__navigate", "arguments": "{\"url\": \"https://google.com\"}"}}
    ]},
    {"role": "tool", "tool_call_id": "call_123", "content": "Navigation complete."}
  ]
}

Statistics

• Total Examples: 1,980
• Train: 1,782
• Validation: 99
• Test: 99

Usage

from datasets import load_dataset

# Load text format
ds = load_dataset("pierretokns/mcp-browser-automation-v2", data_dir="mcp_training_text")

# Load messages format
ds = load_dataset("pierretokns/mcp-browser-automation-v2", data_dir="mcp_training_split")

# Access examples
for example in ds["train"]:
    print(example["text"])  # or example["messages"]

Training

Example training command with MLX-VLM:

uv run python -m src.training.train_vlm_mlx \
  --model mlx-community/LFM2-VL-1.6B-bf16 \
  --data data/mcp_training_text \
  --output models/lfm2-vl-1.6b-ccmcp-v2-lora

License

Apache 2.0

Citation

If you use this dataset, please cite:

@misc{mcp-browser-automation-2026,
  title={MCP Browser Automation Training Dataset},
  author={Pierre Tokens},
  year={2026},
  publisher={Hugging Face},
  url={https://huggingface.co/datasets/pierretokns/mcp-browser-automation-v2}
}