I Downloaded Claude Code’s Leak (The Results are INSANE)
Riley Brown · 2026-04-04
💡 Quick Take
1. Download and experiment with Anthropic's leaked Claude Code source code.
2. Use OpenAI's CodeX to analyze and understand the Claude Code project structure.
3. Run a local version of Claude Code, dubbed "Riley Code," with Anthropic API key.
4. Customize Riley Code by changing action verbs and personality traits.
5. Explore and implement new, funny, and culturally relevant verbs for the agent.
6. Modify Riley Code's personality to be like a direct, fast-paced, smart 19-year-old intern.
7. Develop a desktop application interface for Riley Code, inspired by T3 Chat and OpenAI CodeX.
8. Implement features like workspace files, tool call visualization, and a built-in web browser.
9. Allow for dangerous permission skipping by default in the desktop app.
10. Integrate a web browser within the app that opens automatically with links from responses.
11. Refine the UI/UX of the desktop app, including tool call display, side panel elements, and background customization.
12. Create a PDF file about Riley Code's personality and style it with a green background and black text.
13. Clean up files and structure to resemble OpenAI's CodeX interface.
📊 Detailed Explanation
1. Download and experiment with Anthropic's leaked Claude Code source code. This is the foundational step. The transcript highlights the accidental leak of Claude Code's source code via an npm registry map file. The author's immediate action was to download and explore this leaked code, setting the stage for all subsequent experimentation and modifications. It's about seizing an unexpected opportunity to peek under the hood of a cutting-edge AI project.
2. Use OpenAI's CodeX to analyze and understand the Claude Code project structure. To make sense of the downloaded code, the author turned to OpenAI's CodeX. By uploading the Claude Code project to CodeX and asking it to summarize the agent, they gained a high-level understanding of its components. CodeX identified key elements like the CLI entry point, remote control, daemon workers, background sessions, MCP helpers, and multi-agent delegation, revealing that the leaked code contained the full functionality of Claude Code.
3. Run a local version of Claude Code, dubbed "Riley Code," with Anthropic API key. The ultimate goal was to run the code locally and modify it. The author instructed CodeX to set up the environment for running the code locally, which required an Anthropic API key. This process resulted in the creation of "Riley Code," a personalized version of Claude Code that could be executed from the terminal, demonstrating the feasibility of running and interacting with the leaked code.
4. Customize Riley Code by changing action verbs and personality traits. A significant part of the experimentation involved personalization. The author identified specific "action verbs" within the code (like "spelunking," "working," "symbioting") and expressed a desire to change them to something more aligned with their preferences. This points to the underlying structure of the agent allowing for fine-grained control over its language and behavior.
5. Explore and implement new, funny, and culturally relevant verbs for the agent. To make the customization fun and engaging, the author prompted CodeX to generate a list of 20 humorous and popular cultural verbs (like "jester maxing," "clout maxing," "mewing"). These were then integrated into Riley Code, leading to amusing interactions where the agent used these new terms in its responses, showcasing the flexibility of the code in adopting new linguistic styles.
6. Modify Riley Code's personality to be like a direct, fast-paced, smart 19-year-old intern. Beyond just verbs, the author aimed to alter the agent's overall persona. They requested a personality shift to that of a highly intelligent, direct, and fast-paced 19-year-old intern, communicating in a style reminiscent of a group chat. This involved creating a dedicated `personality.txt` file and modifying the launcher to inject this new persona, demonstrating a more complex level of behavioral customization.
7. Develop a desktop application interface for Riley Code, inspired by T3 Chat and OpenAI CodeX. The ambition extended to creating a user-friendly desktop application. Inspired by the sleek interfaces of T3 Chat and OpenAI's CodeX, the author tasked CodeX with building an Electron app. This app would provide a graphical way to interact with Riley Code, mirroring the terminal experience but with a more polished UI.
8. Implement features like workspace files, tool call visualization, and a built-in web browser. The desktop app development involved incorporating several advanced features. This included implementing workspace files for better project management, visualizing tool calls clearly within the interface, and crucially, adding a built-in web browser. The goal was to create a seamless experience where links clicked within the chat would open directly in this integrated browser.
9. Allow for dangerous permission skipping by default in the desktop app. For the sake of convenience and potentially faster workflow, the author requested the ability to "dangerously skip permissions by default." This implies a setting that bypasses standard security checks, allowing the app to execute actions more freely, though with inherent risks.
10. Integrate a web browser within the app that opens automatically with links from responses. A key UI enhancement was the integrated web browser. The requirement was for any link generated in a response to automatically open within this browser. This also extended to file paths, suggesting a robust system for handling and displaying various types of generated content directly within the application.
11. Refine the UI/UX of the desktop app, including tool call display, side panel elements, and background customization. Significant effort was put into refining the user experience. This involved detailed instructions on how to display tool calls (within a "dot dot dot" animation), adjust side panel elements, change the font size, remove unnecessary lines, and customize the background to match a desired aesthetic (initially inspired by an image, then styled with a green background and black text for a PDF). The aim was to create an intuitive and visually appealing interface.
12. Create a PDF file about Riley Code's personality and style it with a green background and black text. As a test and demonstration of the app's capabilities, the author instructed Riley Code to create a PDF about its own personality. The specific styling request for this PDF (green background, black text, no red, and larger font) highlights the control over output generation and presentation.
13. Clean up files and structure to resemble OpenAI's CodeX interface. The final stages involved tidying up the codebase and interface elements to closely mimic the structure and appearance of OpenAI's CodeX. This included organizing global files, enabling editing of tool files, and creating new sessions, all contributing to a familiar and professional development environment.
🎯 Expert Opinion
This whole situation with the Claude Code leak is a fascinating, albeit slightly chaotic, peek behind the curtain of AI development. It's a stark reminder that even with the best intentions, code can get out there, and the community's ingenuity in responding to it is incredible. The author's approach – downloading, dissecting with another AI (CodeX, which is a smart move to avoid potential data contamination), and then actively modifying – is a textbook example of reverse engineering and creative adaptation in the AI space.
What's particularly interesting is the progression from a raw code dump to a fully functional, personalized desktop application. The ability to not just change verbs but to fundamentally alter the agent's personality and then build a sophisticated UI around it speaks volumes about the modularity and extensibility of these large language models and their surrounding frameworks. The "Riley Code" project effectively demonstrates that these AI systems aren't just black boxes; they have underlying architectures that can be understood, tinkered with, and ultimately, reshaped.
The integration of culturally relevant slang like "jester maxing" and "clout maxing" is a brilliant way to showcase the agent's adaptability and the human element of AI interaction. It moves beyond mere task execution to something more expressive and, frankly, fun. This trend of infusing AI with personality and cultural nuances is a major direction in the field. We're seeing a shift from purely functional assistants to more engaging, relatable digital companions. The author's experiment here is a micro-level example of that macro trend.
The creation of a desktop app inspired by CodeX and T3 Chat is also highly relevant. The demand for polished, integrated development environments for AI tools is growing exponentially. Developers want more than just a chat interface; they want integrated debugging, file management, and visualization tools. The author's success in building this desktop app, complete with a built-in browser and dynamic tool call display, hints at the future of AI development platforms. We're likely to see more such applications emerge, bridging the gap between raw AI capabilities and professional software development workflows.
The "dangerously skip permissions" feature, while risky, highlights a common tension in development: speed versus security. In rapid prototyping and personal experimentation, developers often prioritize getting things done quickly. This is a space where user-friendly interfaces can sometimes abstract away critical security considerations, and it's something that needs careful balancing as these tools become more mainstream.
Finally, the author's iterative refinement of the UI, from the placement of buttons to the transparency of backgrounds and the display of tool calls, is exactly how good user experiences are built. It's about understanding the user's cognitive load and designing an interface that minimizes friction and maximizes clarity. The fact that they could achieve this level of customization through conversational prompts with CodeX and by directly manipulating the code is a testament to the power of AI-assisted development. This entire process serves as a fantastic case study for anyone interested in understanding the practical application, customization, and potential evolution of advanced AI systems.
Kanal: Riley Brown