feat: add AWT (AI Watch Tester) — AI-powered E2E testing MCP server

[ksgisang]

Description

Adds AWT (AI Watch Tester) to the community servers list — alphabetically between AltTester and Amplitude.

What makes AWT different

Most testing MCP servers wrap Playwright for DOM-based testing. AWT goes further:

• Works on Canvas & Flutter Web — uses OpenCV + OCR for visual matching. DOM-based tools fail completely on CanvasKit-rendered apps.
• Self-healing DevQA Loop — test fails? AI analyzes the screenshot diff, patches the selector or scenario, retries automatically (up to 5 attempts).
• Zero test code — point at a URL, AI writes the test scenario. No Playwright code to write or maintain.

The aat_devqa MCP tool

Single tool call that orchestrates the full loop:

scan → generate → run → [fail → AI fix → retry] → report

Installation

pip install aat-devqa
claude mcp add awt -- python mcp/server.py

Links

• GitHub: https://github.com/ksgisang/AI-Watch-Tester
• PyPI: https://pypi.org/project/aat-devqa/
• Demo (30s GIF): https://raw.githubusercontent.com/ksgisang/AI-Watch-Tester/main/docs/assets/demo.gif

[ksgisang]

Added a 30-second demo video to README showing the DevQA Loop in action.

Video: https://github.com/ksgisang/AI-Watch-Tester#readme

[ksgisang]

Demo

Here's a 30-second demo showing install → test execution → self-healing loop:

What makes AWT different from other Playwright/testing MCP servers:

1. Works on Canvas & Flutter Web — uses OpenCV + OCR for visual matching, not just DOM selectors. Most testing tools fail completely on CanvasKit-rendered apps.

2. Self-healing DevQA Loop — when a test fails, AI analyzes the screenshot diff, patches the failing step, and retries automatically (up to 5 attempts). No human intervention needed.

3. Zero test code — give it a URL, it writes the test. Give it a broken test, it fixes itself.

Happy to answer any questions from reviewers.

[olaservo]

Thanks for contributing to the MCP servers repo!

We've retired the third-party server list from the README in favor of the MCP Registry — see #3950 for context. The Official Integrations and Community Servers sections no longer exist in the README, so this PR no longer has a target.

Going forward, please publish your server to the MCP Registry using the quickstart guide:
https://github.com/modelcontextprotocol/registry/blob/main/docs/modelcontextprotocol-io/quickstart.mdx

That's where server discovery now happens. Thanks for building on MCP!

---

Note: I'm working through the backlog with help from Claude Code to give everyone a prompt response.