Codex Just Did What Claude And ChatGPT Couldn’t (Full App, No Code)
The creator demonstrates OpenAI's Codex app, positioning it as a major upgrade over ChatGPT by showing it autonomously build a full web app, design UI screens, catch its own bugs, and generate pitch decks and launch videos — all without writing code. He argues Codex functions as an 'agent' that does work rather than just answering questions, and has switched his entire team to it.
Summary
The video opens with the creator, Vibhav, declaring he has asked his team to stop using ChatGPT in favor of Codex, framing it as the 'full version' of AI that was previously locked behind developer tools. He positions ChatGPT as outdated ('your father's AI') and Codex as the next evolution, powered by GPT-4.5 underneath and now accessible to non-developers via a downloadable Mac/Windows app.
Vibhav walks through the Codex interface, highlighting three core sections: Projects, Plugins, and Automations. He demonstrates how Plugins connect Codex to tools like Gmail, Slack, Notion, Figma, and Canva, enabling it to operate across an entire software ecosystem. His first use case is a daily automation: every morning at 9 a.m., Codex checks his Gmail, finds a specific newsletter, extracts key points, and generates a PowerPoint deck in Canva — entirely without human involvement. He frames this as saving him an hour every day.
The centerpiece of the video is building a full web app called 'Near Folk' — a platform to match people in Bangalore and San Francisco with offline meetups and founder events. Rather than immediately coding, Codex first asks clarifying product questions (user type, data source, matching logic), which Vibhav contrasts favorably against older AI tools that generate random output without seeking clarity. Codex even proposes brand name options, with Vibhav selecting 'Near Folk.'
For UI design, Vibhav uses a tool called 'Paper' connected to Codex via MCP (Model Context Protocol), allowing Codex to autonomously control Paper's canvas and draw app screens in real time — hero sections, buttons, chat interfaces, and marketplace layouts — without any human input on the design tool itself. He describes this as a brain-breaking moment: an AI driving another piece of software the way a human would.
In 'Plan Mode,' Codex maps out the full build architecture — database structure, page routes, and components — before writing a single line of code. During the build, Vibhav uses a 'Steer' feature to adjust scope mid-build without interrupting the agent. He also demonstrates leaving plain-English comments in the code (to switch from OpenAI API to OpenRouter) rather than editing code directly, showing that code literacy is not required.
The most dramatic moment Vibhav highlights is Codex autonomously catching and fixing a mobile layout bug during its own self-testing phase — no human prompt required. It identified the issue, planned a fix, implemented it, and verified the result in roughly 20 seconds while Vibhav was in the kitchen. He notes this kind of autonomous QA is something he normally has to push technical teams to do for weeks.
Using a 'Fork into Local' feature, Vibhav runs a parallel chat thread to generate launch materials simultaneously — a four-to-five slide sponsor pitch deck and an 8-second branded launch video with music and motion graphics — from a single prompt. The video closes with a live demo of the completed Near Folk app, showing functional event browsing and seat-request forms for both Bangalore and San Francisco.
Key Insights
- Vibhav argues that Codex is not just a better chatbot but a fundamentally different category of tool — an 'agent' that does things (opens files, runs software, sends Slack messages, builds apps) rather than merely answering questions, which he identifies as the core distinction from ChatGPT.
- Vibhav claims Codex behaves like a product manager rather than a code generator — before building the Near Folk app, it returned clarifying questions about the user, data source, and matching logic, which he contrasts with older AI tools that 'spit out something random' from a single prompt.
- Vibhav describes watching Codex autonomously control the Paper design tool in real time — moving the cursor, drawing UI screens, placing buttons and text — and argues this signals that 'the skill is not the design tool anymore, the skill is knowing what to ask.'
- Vibhav highlights that Codex caught its own mobile layout bug during self-testing — without any human prompt — planned a fix, implemented it, and verified the result in about 20 seconds, calling it 'not a tool, that is a teammate' and noting this level of autonomous QA normally takes weeks to get from technical teams.
- Vibhav demonstrates that plain-English comments left directly in the code — rather than actual code edits — are sufficient for Codex to understand and implement changes, using this to argue that users don't need programming knowledge, only the ability to be specific about what they want.
Topics
Full transcript available for MurmurCast members
Sign Up to Access