Riley Brown·April 4, 2026·17m

Claude Code Was Leaked… So I Had to Test it

TL;DR

Anthropic accidentally exposed Claude Code via an npm map file — Riley Brown grabbed the leaked source before the link stopped working, then used OpenAI Codex to inspect what he says is Anthropic’s full Claude Code CLI with daemon workers, MCP helpers, plugin agents, and multi-agent delegation.
He immediately forked it into “Riley Code” and proved the interface is highly malleable — with a few changes, he renamed the CLI, swapped Claude Code’s status verbs like “symbioting” for meme-y ones like “status maxing,” “clout maxing,” “mewing,” and “aura farming,” and ran it locally from terminal.
The bigger hack was personality injection, not just UI polish — Codex added a dedicated personality.txt prompt file and launcher hook so the agent would reply in lowercase, in short bursts, like a “really smart 19-year-old intern in a group chat.”
Riley then pushed the leaked CLI into a desktop app modeled on Codex/T3 Chat — he had the agent build an Electron wrapper with chat history, tool-call visibility, project/session management ideas, and eventually an in-app browser panel inspired by Claude Artifacts.
The video is basically a live demo of “vibe-coding the vibe-coder” — he uses one coding agent to analyze and reshape another, asking for everything from default dangerously skip permissions behavior to persistent memory files like soul.md and a Blueberry-style visual redesign.
By the end, he has a functioning custom coding app that feels like his own Codex clone running Claude underneath — after another two hours of cleanup, he shows fullscreen panes, editable global files, side-by-side previews, and easy new-project creation, all wrapped around his personalized “Riley Code.”

The Breakdown

The leak that kicked the whole thing off

Riley opens with the kind of sentence that tells you exactly where this is going: Anthropic “made a huge mistake” and apparently leaked Claude Code again through a map file on npm. A download link spread on Twitter, he grabbed the source before it got blocked, and frames the whole stunt with repeated “for entertainment purposes only” and a joking “Anthropic, please don’t hurt me.”

Using OpenAI Codex to dissect Claude Code

Instead of poking around manually, he drops the downloaded project into OpenAI Codex and asks for a summary of the agent. Codex reports that it looks like Anthropic’s Claude Code CLI source tree, with an entry point that prints “Claude Code” plus fast paths for remote control, daemon workers, background sessions, MCP helpers, built-in plugin agents, and even multi-agent delegation — which is the moment Riley gets visibly excited.

From Claude Code to “Riley Code”

His first move is not some deep architectural rewrite — it’s branding and UX. He asks Codex to make the local version runnable as “Riley Code,” wires in his Anthropic API key, and gets a working terminal command so he can launch the forked agent directly instead of the official CLI.

Swapping Claude’s verbs for internet-brainrot status messages

Once Riley Code is running, he goes hunting for the status verbs in the codebase — things like “spelunking,” “working,” and “symbioting.” Then he does the most Riley Brown thing possible: asks for a list of funny internet-culture verbs and replaces them with phrases like “status maxing,” “clout maxing,” “mewing,” and “aura farming,” turning the coding agent into a meme machine without changing its core function.

Personality modding through a prompt file

Next he tries something more interesting than cosmetics: changing the agent’s actual voice. He asks for a persona that responds in all lowercase, in quick bursts, like a sharp 19-year-old generalist intern in a group chat, and Codex implements it by creating a personality.txt file and injecting that persona through the launcher using a supported system prompt path.

Building an Electron app around the leaked CLI

After proving the terminal fork works, he pivots to making a desktop experience that feels more like Codex or T3 Chat. He asks the agent to build an Electron app that exposes the same terminal-style tool calls in a cleaner interface, then keeps layering requests: previous chat sessions, visible tool usage, dangerously-skip-permissions behavior, and styling inspired by T3 Chat and the Blueberry app.

Design nitpicks, memory, and an artifacts-style side browser

This middle stretch is basically live product direction. Riley dictates UI tweaks in real time — move tool status into the three-dot animation, remove “reply ready,” let him type while the agent is still running, simplify the sidebar, add project-based directories and persistent memory files like soul.md, and finally bolt on a right-side browser pane so links, localhost apps, PDFs, and generated files open inside the app like Claude Artifacts.

The next-day reveal: a personalized Codex-style wrapper for Claude

By the end — after another two hours of cleanup the next day — he shows a much more polished version with fullscreen views, side panels, resizable windows, editable global files, and easy new-session creation. His final point is less legal drama than creative possibility: he used one coding agent to reverse-engineer and reshape another, ending up with a custom Codex-like app that runs his own personality-tuned version of Claude Code underneath.