I made Claude Code, Codex, and Antigravity build their own rivals, and one was in a different league

Published Nov 3, 2026, 6:00 AM EST Mahnoor Faisal is a tech journalist covering AI and productivity tools with bylines at XDA, SlashGear, MakeUseOf, Laptop Mag, and Android Police. She's been writing professionally since she was sixteen, and has since penned hundreds of articles. This includes in-depth coverage of AI tools like NotebookLM to breaking news across the AI space. Her passion for technology started when she received her first iPod Touch (4th generation) on her 8th birthday, and she's been deep in the tech world ever since. Currently pursuing a degree in computer science, Mahnoor brings both a journalist's eye and a technical foundation to her coverage of how AI is reshaping the way we work and learn. I love unhinged 3 AM ideas. I love the adrenaline rush that comes with convincing myself that a ridiculous idea might actually work, and then immediately putting it to the test instead of, you know, going to sleep. One of those ideas just so happened to be asking Claude Code to build its own replacement. While that's the sort of experiment I might once have dismissed entirely, the fact that the Claude Code team has previously shared that Claude wrote approximately 90% of the tool's own code gave me a little more confidence. If it could help build itself, surely it could build a lightweight alternative for me, right? As it turned out, it could. Naturally, that left me wondering whether Codex and Antigravity could pull off the same thing. So, I decided to give them their turn. I used the same approach for all three coding agents I gave them all the same existential crisis While you can read a more in-depth breakdown of how I built my Claude Code replacement, XDACode, in a separate article, I used the same approach to build the two new alternatives: XDAGravity and XDACodex. Of course, instead of asking Claude Code to build these, I asked Antigravity and Codex to build their own respective replacements. I gave both agents a detailed prompt outlining the same core requirements. To make it a fair test, I used the same prompt I used to build the Claude Code replacement. The goal here wasn't to make a feature-by-feature replica of each tool. All I wanted was a functional coding agent that could inspect a project, read and modify files, execute commands, run tests, and fix errors on its own. I also made it clear that they shouldn't simply build wrappers around existing coding agents or rely on a coding-agent SDK to do all the heavy lifting. Beyond that, I left the technical decisions up to them, including which programming language to use, how to structure the application, and which libraries to install. Claude Code and Codex took a similar route Antigravity went its own way Since I'd left the technical decisions up to each coding agent, I was curious to see how differently they'd approach the same task. Claude Code and Codex ended up making a similar first decision: both chose Python to build their replacements. Claude Code used the official Anthropic SDK to communicate with Claude's models, along with Rich and prompt_toolkit for the terminal interface. Codex, on the other hand, built its initial version using Python's standard library, without any runtime dependencies, and connected directly to OpenAI's API. Both built their own agentic loops to coordinate file operations, execute commands, inspect results, and keep working through a task. Antigravity took a different approach. It chose Node.js and also built its initial version without external dependencies. What caught my attention more, though, was that it created two ways to use its coding agent: an interactive terminal interface and a browser-based workbench. The browser interface included a project file explorer, a live conversation and tool-execution feed, approval prompts, and a panel for reviewing changes. It also offered the option to revert individual files or undo the entire session's changes. Meanwhile, the terminal version came with slash commands for switching models and providers, reviewing diffs, and reverting files. That made Antigravity's implementation the most ambitious of the three in terms of interface. Claude Code and Codex had focused on delivering terminal-based coding agents, while Antigravity had given me a visual workspace alongside one. Codex has been one of my go-to coding agents lately, and despite taking the longest of the three to build its own replacement, it really disappointed me visually. When a coding agent is working, I don't like seeing every single line of code it writes or changes filling up my terminal. I'd much rather see a clear summary of what it's doing, with the option to inspect the code when I actually want to. XDACodex's output felt cluttered, especially compared to XDACode and XDAGravity, which kept things simpler by showing the file being written and the number of lines instead! I added ChatGPT sign-in, then put all three replacements to the test One subscription to power them all Similar to how I had done with the Claude Code test, the next thing I wanted to do was get both replacements working with my existing ChatGPT Pro subscription. I didn't want to keep paying separately for API credits when I already had a subscription I could use. So, I gave both agents a follow-up prompt asking them to add OpenAI's Sign in with ChatGPT integration, along with links to the official documentation. Both managed to add the integration, and it didn't take long for me to get it set up and working. Though Codex did take a bit longer, it spent more time testing the authentication flow and handling potential errors. Antigravity needed a little troubleshooting after sign-in because it initially selected an unsupported model, but once that was sorted, I could move on to testing the agents themselves. It was then time to put all three to the test. Now, given that all three were using the same models under the hood, there wasn't really a point in comparing the results purely on how good the generated code or app looked. I was more interested in how each agent handled the process: whether it could create and edit files, run commands, test its work, and recover when something went wrong. XDACode takes first place, with XDAGravity ahead of XDACodex Claude Code’s replacement won me over on experience alone Given that the output each tool produced wasn't my main focus here, I judged them primarily on the user experience. I wanted to see which replacement felt the most comfortable to use, how clearly it communicated what it was doing, and whether the interface made it easy to follow along. On those fronts, XDACode was my favorite, followed by XDAGravity and then XDACodex. XDACode struck the best balance between keeping me informed and keeping the terminal clean. Even smaller details impressed me, like the subtle "planning..." or “preparing write_file…” indicator that appeared while it prepared a file. Once it was ready, it displayed the filename and line count in a neatly boxed summary and asked whether I wanted to apply the change. I knew what was happening without having to scroll through every line of code. XDAGravity came second. Its use of color made it easy to distinguish responses, progress updates, and completion messages. I also appreciated having commands for switching models, reviewing changes, and reverting files within the session. The repeated “TASK SUCCEEDED” messages felt unnecessary for something as simple as saying hello, but the interface was still easy enough to follow. Like mentioned above, XDACodex disappointed me here. It printed lengthy code diffs directly into the conversation, which quickly made the terminal feel cluttered. Some people might appreciate that level of visibility, but I'd rather have a concise summary and inspect the changes separately when I need to. It also displayed an “[UNVERIFIED]” warning after responding to a simple greeting because the model hadn't called its completion tool! Nothing had actually gone wrong with the greeting, yet the message made it seem like something had. Combined with the repeated authentication details at startup and the lack of visual separation between different types of output, it made XDACodex feel rougher around the edges. Of course, these are things I could ask the agents to improve with a few follow-up prompts. But judging the versions they delivered, XDACode was the one I most enjoyed using. Ironically, XDACodex was the quickest to call it a day What made me laugh was that Claude Code was the quickest to finish building XDACode, followed by Antigravity building XDAGravity, while Codex took the longest to build XDACodex. Somehow, the tool that spent the most time building its replacement ended up creating the one I liked the least!