PipelineLabs/WebEnvDemo

Web App Environments and Cross-App RL Tasks Demo

This demo features three web applications—SiYuan for note-taking, SignaturePDF for PDF editing, and Docmost for online document collaboration—along with some samples of cross-app RL tasks (easy-level) designed by our team. That is, to finish each RL task, the agent must interact with multiple App instead of single one. Each task comes with accurate outcome reward to guild training and fine-grained evaluation.

We evaluate these tasks using frontier LLMs under the MCPMark framework, where models interact with the apps through Playwright MCP to complete the tasks. The setup can also be readily extended to other agent harnesses, such as GUI Agent and Browser-Use agents.

Installation

Create a conda environment with the required system libraries (needed for Playwright/Chromium on headless servers):

cd webenv_demo
conda create -n webenv python=3.11 -y
conda activate webenv
pip install -e .
playwright install

Quick Start

1. Configure Server Hostname & API Keys

Run the one-click setup script:

./start.sh

This will:

• Replace the <your-server-hostname> placeholder in all task files with your the app-deployed server's IP or just localhost.
• Configure OpenRouter API proxy keys in .mcp_env

To manually configure API keys (OpenAI, Anthropic, Google, etc.), edit .mcp_env directly.

2. Deploy Docker Environments

Docker images are placed in the docker/ directory.

cd docker

# SiYuan (note-taking)
docker load -i mw-siyuan.tar
docker run --name siyuan -p 6037:6806 -d mw-siyuan:latest

# SignaturePDF (PDF editing)
docker load -i mw-signaturepdf.tar
docker run --name signaturepdf -p 6093:80 -d mw-signaturepdf:latest

# Docmost (wiki/docs)
docker load -i mw-docmost.tar
docker run --name docmost -p 6001:3000 -d mw-docmost:latest

After a few minutes, verify the services are running:

• SiYuan: http://<your-server-hostname>:6037
• SignaturePDF: http://<your-server-hostname>:6093
• Docmost: http://<your-server-hostname>:6001

3. Run Evaluation (with Gemini-3-Flash)

Make sure the webenv conda environment is active (conda activate webenv), then:

# All tasks
python -m pipeline --mcp playwright_webarena --models openai/google/gemini-3-flash-preview \
    --task-suite standard --exp-name test_all --k 1

After the evaluation, you MUST run the reset.sh to reset the environment.

4. Aggregate Results

python -m src.aggregators.aggregate_results --exp-name <your-exp-name>

CLI Reference

Argument	Default	Description
--mcp	filesystem	MCP service to use
--models	(required)	Comma-separated model list
--agent	mcpmark	Agent implementation
--tasks	all	"all", "category", or "category/task_id"
--task-suite	standard	standard or easy
--exp-name	timestamp	Experiment name (required when --k > 1)
--k	4	Number of runs for pass@k metrics
--timeout	3600	Agent execution timeout (seconds)
--reasoning-effort	default	minimal / low / medium / high / default

License

See LICENSE.