Reading the leaked Claude Code source through an MCP server

What the repository contains and how it got there

Claude Code is Anthropic's command line tool for working with Claude from a terminal: editing files, running commands, searching codebases and driving git workflows. This repository holds a copy of its src/ directory.

The README states that Chaofan Shou, posting as @Fried_rice, found that the published npm package for Claude Code shipped a .map file pointing at the full, unobfuscated TypeScript source, downloadable as a zip from Anthropic's R2 storage bucket. The leak is dated 2026-03-31 in the README, and the quote and link to the original post are reproduced in it. The repository keeps the original unmodified leaked source on a branch named backup, which is the place to go if you want the material as it arrived rather than the annotated version on the default branch.

Worth saying plainly: this is a third party mirror of source that was not published on purpose. Nothing here is an official release channel.

The MCP server is the useful part

The repository ships a Model Context Protocol server so any MCP compatible client can query the source instead of opening files one at a time. The README names Claude Code, Claude Desktop, VS Code Copilot and Cursor as clients.

Eight tools are documented. list_tools enumerates the agent tools with their source files, list_commands does the same for slash commands, get_tool_source returns the full source of one tool with BashTool and FileEditTool given as examples, and get_command_source does the same for a command such as review or mcp. read_source_file reads any file under src/ by path, search_source greps the whole tree, list_directory browses it, and get_architecture returns a high level overview.

Four prompts sit alongside them: explain_tool, explain_command, architecture_overview and how_does_it_work, the last of which takes a subsystem name such as permissions, MCP or bridge. In practice the pair that changes how you read a large unfamiliar codebase is search_source plus read_source_file, since it turns a grep session into a question you ask an assistant.

Installing the explorer from npm or from source

The server is published on npm as warrioraashuu-codemaster, so the shortest path needs no clone. This registers it with Claude Code:

claude mcp add warrioraashuu-codemaster -- npx -y warrioraashuu-codemaster

If you want the local build instead, clone the repository, build the server under mcp-server, and register the compiled entry point. The README gives it as a one-liner:

git clone https://github.com/codeaashu/claude-code.git ~/claude-code \
  && cd ~/claude-code/mcp-server \
  && npm install && npm run build \
  && claude mcp add claude-code-explorer -- node ~/claude-code/mcp-server/dist/index.js

After that, asking the client to list tools or grep the source should route to the server. The README notes that npm run build has to run before registration, since the registered path is dist/index.js.

For VS Code, the README says to add this to .vscode/mcp.json, where the environment variable tells the server where the source root lives:

{
  "servers": {
    "claude-code-explorer": {
      "type": "stdio",
      "command": "node",
      "args": ["${workspaceFolder}/mcp-server/dist/index.js"],
      "env": { "CLAUDE_CODE_SRC_ROOT": "${workspaceFolder}/src" }
    }
  }
}

Claude Desktop uses the mcpServers shape instead, with absolute paths, and Cursor uses the same format in ~/.cursor/mcp.json.

What the docs directory covers

Five guides live under docs/. architecture.md covers the core pipeline, startup sequence, state management, rendering and data flow. tools.md is a catalog of the agent tools with categories and the permission model. commands.md covers the slash commands by category. subsystems.md goes into the Bridge, MCP, permissions, plugins, skills, tasks, memory and voice subsystems. exploration-guide.md is the one to read first, since it gives study paths, grep patterns and the key files.

One inconsistency is worth flagging before you rely on counts. The docs table describes commands.md as a reference for about 85 slash commands, while the MCP tool table describes list_commands as listing about 50. Both numbers come from the same README. Either one of the two is stale, or they count different things, and the repository does not say which.

Scale and stack of the leaked tree

The README puts the source at roughly 1,900 files and more than 512,000 lines of TypeScript, written in strict mode. The runtime is Bun, which matters if you intend to run any of it rather than read it, and the terminal UI is React rendered through Ink.

The repository listing around that source includes docs/, prompts/, mcp-server/, docker/, web/, scripts/, plus Skill.md and agent.md at the root, a Dockerfile, vercel.json, server.json and a gitpretty-apply.sh script. The npm badge on the README points at warrioraashuu-codemaster, not at the repository name, which is easy to trip over when you are looking for the package.

Provenance, licence and why they matter here

GitHub reports the licence as NOASSERTION even though the listing includes a LICENSE file. That combination means whatever text is in that file was not matched to a recognised licence, and the source itself belongs to Anthropic. Reading it to learn how agent tooling is built is one thing; copying it into a product is a different question that this review will not answer.

The second problem is time. This is a snapshot dated 2026-03-31, and the tool it came from ships continuously. Any behaviour you deduce from this tree may already be wrong upstream, and there is no version tag or release in the repository to anchor it to.

A smaller note on the README itself: it carries a promotional block for an unrelated service, RepoXray, in the middle of the page. It does not affect the source, but it is a signal about how the repository is maintained. The last push was on 2026-08-29.

Learning the same patterns without the leak

The alternative is to study Anthropic's official Claude Code documentation and release notes instead of this tree. The difference is one of layer. Official material describes the interface: what commands exist, what permissions do, how configuration is read. It tells you nothing about how those features are implemented, which is exactly the layer this repository exposes.

If the implementation layer is what you want, the honest options are this snapshot, with its licensing and staleness problems, or building your own agent tooling and hitting the same design problems yourself. What this repository is genuinely good for is answering narrow questions about structure, such as how the permission prompt is wired or how the bridge talks to an IDE, where reading a real implementation beats reading a changelog.

Editorial conclusion

Read this repository if you want to see how a mature terminal coding agent is put together: the permission model, the tool catalog, the bridge to editors. Do not build on it, given the NOASSERTION licence on someone else's source and the fact that it is a 2026-03-31 snapshot of a product that keeps shipping. Start with docs/exploration-guide.md, and if you only want to query the source, install the npm package and skip the clone.

Frequently asked questions

Official sources

1. codeaashu/claude-code on GitHub
2. Issues
3. README