Austin1serb/Anthropic-Leaked-Source-Code: what is actually in the archive

What the Claude Code leak archive actually contains

This repository is an archive of source code attributed to Claude Code, Anthropic's AI coding agent. The README describes it as unmodified: "Nothing has been added besides this README. Nothing has been removed or changed." The top-level listing backs that claim in structure if not in content, with entry points such as main.tsx, QueryEngine.ts, Task.ts and Tool.ts sitting alongside directories for tools, services, skills, plugins, schemas, migrations and a vim mode.

The audience named in the README is developers, researchers and security-minded engineers who want to study how a modern AI coding agent is structured. That framing matters. Nothing in the repository is presented as a build of the product, and the README does not describe a way to run the agent. You are being handed a directory tree and asked to read it.

The topics list is the other half of the picture: anthropic-leaked, claude-code-leaked, code-leak, source-code-leak. The project positions itself as an archive of a leak, and the README repeats the search vocabulary people use for it ("Claude Code leak", "Anthropic source code leak", "Claude Code source map"). That is a discoverability decision, not a technical one.

How the source maps became TypeScript

The README credits Serbyte Web Design & Development with reverse engineering the code "using JavaScript source maps". That single sentence is the most important technical fact about the archive, because it explains both what you get and what you cannot trust.

Source maps exist to map minified or transpiled JavaScript back to original positions. When the original sources are embedded, the map carries them; when they are not, tooling reconstructs file paths, names and approximate structure from mapping entries. The result can look like a coherent TypeScript project while remaining a reconstruction. The README does not say which case applies here, whether every file in the listing was recovered the same way, or how much of the type information survived. A reader should assume that type annotations, generics and imports are the least reliable parts of the tree until proven otherwise.

The architecture visible in the listing is the part that survives reconstruction reasonably well, because directory and file names are structural. There is a query layer (QueryEngine.ts, query/, query.ts), a task layer (Task.ts, tasks.ts, tasks/), a tool layer (Tool.ts, tools.ts, tools/), a service layer (services/, server/), a transport layer (bridge/, remote/, upstreamproxy/), plus schemas, state, hooks, keybindings and output styles. That is a conventional shape for a terminal agent, and it is the shape most readers come for.

Reading the tool, permission and MCP layers without running them

The README lists what is included: agent orchestration and query engine, tool implementations for Bash, FileEdit, FileRead, FileWrite, Glob, Grep, WebSearch, WebFetch and Notebook, an API service layer with model provider integrations, MCP (Model Context Protocol) server management, a permission system with sandboxing logic, CLI transports including SSE, WebSocket and hybrid, and prompt construction with system prompt logic.

For an engineer evaluating agent design, the interesting question is where the trust boundary sits. The presence of both a permission system and sandboxing logic in the same archive, separated from the tool implementations, suggests the tools are meant to declare what they need and something else decides whether they may have it. The README does not document that interface, and the archive alone will not tell you whether the check happens before or after a tool call is dispatched.

The MCP piece is the other one worth your time. MCP server management appearing as its own concern, rather than folded into the tool layer, is the kind of decision that is easy to describe and hard to get right. Reading the code tells you how this implementation ordered it; it does not tell you whether that ordering was correct, and there is no test suite mentioned in the README to settle the question.

There is no install, so here is how to read it

The README gives no install steps, no package name, no build command and no runtime requirement. It says the repository is an archive for research and education. There is no package.json mentioned in the top-level listing either, which means the usual npm install path is not documented and should not be assumed.

The README does point to one external location. It links to github.com/anthropics/claude-code as the project this archive is attributed to, and it names serbyte.net as the firm that reverse engineered it. Those are the only destinations the README gives.

Since no commands are documented, the useful first move is to open the repository on GitHub and read the top-level listing against the README's "What's included" list. Check whether the entries the README claims, QueryEngine.ts, Task.ts, Tool.ts, tools/, services/, skills/, plugins/, schemas/ and migrations/, are actually present. Then open QueryEngine.ts and read the first screen of it. What you see there, named functions and typed signatures or generated-looking code, tells you more about the archive's usefulness than any description of it. The README does not promise that any of this type-checks.

Where this archive is the wrong tool

The most concrete limitation is legal, not technical: the repository listing shows no licence file, and the README states "No affiliation with Anthropic AI. Claude Code is a product of Anthropic AI." An archive of leaked proprietary source with no licence grant is not something you can copy into your own codebase. The README's own framing, educational and research purposes only, is a stated intent rather than a permission.

The second limitation is that the archive is frozen. The last push was on 2026-06-13, and the README describes the contents as the raw source as leaked. Whatever the upstream product has changed since then is not reflected here. If you want to understand how the agent behaves today, this is the wrong source.

The third is that reconstruction and fidelity are not the same thing. If you are trying to answer a question that depends on exact behaviour, such as the precise order of permission checks or the default timeout on a tool call, a source-map reconstruction is a weak witness. The README does not document rollback, versioning or any verification of the recovered files against an original, so there is no way to distinguish a faithful copy from a plausible one by reading the README alone.

What to compare it against instead

The honest alternative for most readers is Anthropic's own published Claude Code repository, which the README links to at github.com/anthropics/claude-code. The difference in approach is not subtle: that is the vendor's own repository, so what it contains is whatever Anthropic chose to publish, under whatever terms Anthropic attached. This archive is a third-party reconstruction of code the vendor did not publish, under no terms at all.

If your goal is to build an agent, the vendor repository and its documentation are the sources you can act on. If your goal is to study architecture that the vendor has not documented, the archive is the only thing of its kind here, and you accept the fidelity caveats above.

A second comparison is against open agent frameworks generally. Those give you a working runtime, tests and a licence, at the cost of not being this particular design. The archive gives you this particular design and nothing you can run. Pick based on whether you need a system or a specimen.

Maintenance cost and the missing licence

There are no releases, so there is no upgrade path to plan for. The last push was on 2026-06-13, and the README describes the archive as complete and unchanged. Practically, that means the cost of adopting this repository is zero ongoing maintenance and a one-time reading effort, which is the opposite of the usual dependency trade-off.

The licence situation is the part that should stop a commercial reader. The repository listing contains no LICENSE file, and nothing in the README grants rights beyond a stated educational and research purpose. Absent an explicit grant, the default is that the author retains rights, and the underlying code is attributed to Anthropic with an explicit no-affiliation note. I am not a lawyer and this is not legal advice; if you intend to do anything beyond reading, that is a question for someone qualified, and the answer will depend on facts this repository does not supply.

For internal study, the cost is the time to read TypeScript that may not type-check. For anything downstream, the cost is unknowable from what the repository states.

Editorial conclusion

Read this repository if you want to study how a CLI coding agent splits orchestration, tools, permissions and transports, and treat it as a static artefact rather than a dependency. Do not clone it into a product, publish it, or ship anything built from it: there is no licence file, no release, and the README states no affiliation with Anthropic. Before you spend time on it, verify three things yourself: whether a LICENSE file exists at the top level (the listing shows none), whether the TypeScript under QueryEngine.ts, tools/ and services/ is complete enough to type-check, and whether the source-map reconstruction left readable names or minified identifiers in the parts you care about.

Frequently asked questions

Official sources

1. Austin1serb/Anthropic-Leaked-Source-Code on GitHub
2. Issues
3. Project website
4. README