AI Tester for Vibe Coding: E2E Framework Connected to Claude and Codex

⚡ The gist in 5 seconds

• An E2E framework for web and mobile connects to Claude, Codex, and other AI models as a tester.
• A single API covers iOS, Android, and the browser at once — the agent doesn't need to write a separate runner for each target.
• Tests are deterministic: steps are fixed, results are reproducible, not "the model clicked wherever it happened to."

🔍 What happened

The GitHub Telegram channel showed an e2e framework for testing web and mobile apps that connects to Claude, Codex, and other AI models. The task is simple on paper: an AI agent writes code — and that same agent, or a separate model, runs it through real scenarios in the app. One API for three platforms: iOS, Android, and web.

The mechanics, broken down step by step. The agent formulates test steps, and the framework turns them into deterministic actions in the app — a tap, input, navigation, a screen-state check. Determinism is key here: the same scenario must produce the same result on repeat, otherwise the test is useless. The post's author emphasizes that the agent "won't have to reinvent" the plumbing for each platform every time — a unified interface already exists.

🧠 Our take

Vibe coding has hit a boring bottleneck. A model can sketch out a screen and a handler in a minute, but verifying that the button actually presses and the form actually submits is still painful: UI tests are the most fragile layer of development — they break over a timing change or a single extra pixel. Our assessment: demand for an "AI tester" grew not because new models appeared, but because there are now lots of agents and no way to check their work.

What this changes in practice. If the framework really provides a stable bridge between LLMs and native drivers, then the "generate — run — fix" loop closes without a human in the middle. That's no longer autocomplete — it's a different mode of working. From experience, this is exactly where most "autonomous" pipelines fall apart: code gets generated fast, verification is done by hand.

Who wins. Teams without a dedicated QA person for mobile — and there are plenty of those. Who's under pressure: those selling manual plumbing — classic UI runners where iOS and Android still require different configs. But my bet is that the winners won't be "replace QA" but "remove the routine": deterministic tests will still need human review, otherwise the agent will show green where the app is actually broken.

⚠️ What we don't know

The post gives no project name, version, license, or repo link — just a description. It's unknown whether the tests run on real devices or emulators, how many platforms are actually covered rather than claimed, and who the author is. A single Telegram post is weak confirmation: this could be a working open-source project or an early prototype. You have to check it yourself: download it, run the same scenario ten times, and see if there's any flakiness.

Prediction: within a quarter, at least two public competitors with MCP-compatible integration will emerge around AI agents for e2e — the market for agent testing plumbing is just opening up.

Disclaimer: one source, no official announcement with a project name or version — details unconfirmed.

Source: t.me/github