Skip to content

Legal and Ethics

This is an independently written research and documentation project. It is not affiliated with, sponsored by, or endorsed by Anthropic. Nothing on this page is legal advice.

Applicable product terms

Anthropic’s Claude Code legal and compliance page says use is governed by:

At the time this atlas was prepared, Commercial Terms section D.4 prohibited customers from reverse engineering or duplicating the Services. Consumer Terms section 3 prohibited decompiling, reverse engineering, disassembling, or otherwise reducing the Services to human-readable form, except where such restrictions are prohibited by applicable law.

These provisions create material contractual risk for reverse engineering and especially for publishing recovered source. Ownership of a local copy does not by itself grant redistribution rights. Researchers should obtain written permission or qualified legal advice for their jurisdiction and agreement before proceeding beyond the narrow public approach used here.

Publication boundary

This repository publishes:

  • independently authored prose, diagrams, contracts, and state machines;
  • cryptographic hashes, sizes, offsets, virtual module names, and signature metadata;
  • short functional identifiers needed to identify interfaces and security controls;
  • read-only CLI help generated by the installed tool;
  • reproducible tooling based on public Mach-O and MIT-licensed Bun formats.

It does not publish:

  • the Claude Code executable;
  • extracted JavaScript or bytecode;
  • native add-on bytes;
  • deminified or reformatted proprietary function bodies;
  • source maps, if later found;
  • Anthropic logos, visual trade dress, or marketing assets;
  • credentials, settings, transcripts, memory, debug logs, or user data.

The local extraction feature writes to ignored working storage. A contributor must never move that output into docs/, evidence/, reconstructed/, issues, pull requests, or release artifacts.

Licensing

Any repository license applies only to original project documentation, inspection tooling, and independently authored reconstruction. It does not license Anthropic software, Claude/Anthropic trademarks, third-party embedded dependencies, or any material excluded above.

Open-source licenses for components such as Bun do not convert the combined proprietary application into open source. Third-party attribution and Anthropic application rights must be evaluated separately.

Trademark and identity

“Anthropic,” “Claude,” and “Claude Code” identify the product and owner being discussed. The project should:

  • use plain-text nominative references only;
  • avoid Anthropic logos, product icons, and confusingly similar branding;
  • keep the independent-project notice visible on the site and repository;
  • avoid names or descriptions implying an official source release;
  • never claim compatibility, endorsement, or completeness without evidence.

Responsible security research

Anthropic’s Responsible Disclosure Policy requests prompt private reports through its HackerOne path, minimal exploitation, no harm or unrelated data access, no retention of inadvertently accessed data, and coordination before public disclosure. Questions about permitted research can be directed to disclosure@anthropic.com before testing.

The policy’s safe harbor is conditional and ultimately determined by Anthropic. It also excludes third-party-controlled systems. Researchers must comply with applicable law and the relevant third party’s disclosure process.

Privacy and data minimization

Use synthetic repositories, dedicated test credentials, and a clean home/config directory. Do not inspect another user’s account, Keychain, sessions, local IPC, browser, IDE, or MCP servers. Do not publish absolute personal paths when $HOME is sufficient. Revoke any credential that appears in a trace rather than relying on redaction alone.

Contribution rules

Contributors must certify that a contribution:

  1. is independently authored or lawfully sourced;
  2. contains no leaked, employee-only, or confidential Anthropic material;
  3. contains no copied recovered implementation body;
  4. contains no user/customer data or live secret;
  5. labels observation, derivation, and hypothesis accurately;
  6. does not disclose an uncoordinated vulnerability or weaponizable exploit detail.

Maintainers may remove material immediately if provenance or disclosure status is uncertain.

Corrections and takedown

Technical corrections should include the artifact hash and contrary evidence. Legal, privacy, trademark, or disclosure concerns should be handled privately first. If sensitive material reaches Git history or Pages, disable publication, remove the material from history where appropriate, rotate affected secrets, preserve minimal private incident evidence, and notify the proper owner.