United States Code Available as Git Repository with Full Change History

The United States Code is now available as a Git repository containing every title, chapter, and section of federal law stored as Markdown files. Each commit corresponds to an Office of the Law Revision Counsel (OLRC) release point, providing point-in-time snapshots of federal law from 2013 to present.
Repository Structure and Contents
The repository contains 53 titles of the United States Code with approximately 2,950 chapter-level Markdown files and 60,400 sections with full statutory text. The data is organized in a directory structure like:
uscode/ ├── title-01-general-provisions/ │ ├── _title.md │ ├── chapter-001-rules-of-construction.md │ └── ... ├── title-02-the-congress/ ├── title-03-the-president/ └── ...
Each Markdown file includes YAML frontmatter with metadata (title number, chapter, heading, section count, source URL) and the full statutory text with cross-references, statutory notes, and amendment histories.
Git Commands for Legal Analysis
Developers can use standard Git commands to analyze legal changes:
git diff congress/115..congress/116 --stat- See what changed between the 115th and 116th Congressgit diff annual/2019..annual/2025 -- uscode/title-18-crimes-and-criminal-procedure/- Full text diff of Title 18 between 2019 and 2025git log --oneline -- "uscode/title-18-crimes-and-criminal-procedure/chapter-044-firearms.md"- See when a specific section was last modifiedgit diff annual/2013..annual/2025 --stat | tail -1- Count files changed across the 12-year history
Tags and Historical Points
The repository includes tags for specific Congress sessions and years:
annual/2013- 113-21 (2013, earliest available OLRC snapshot)congress/113- 113-296 (2014, end of 113th Congress)annual/2015- 114-38 (2015)congress/114- 114-329 (2017, end of 114th Congress)annual/2017- 115-51 (2017)congress/115- 115-442 (2019, end of 115th Congress)annual/2019- 116-91 (2019)congress/116- 116-344 (2021, end of 116th Congress)annual/2021- 117-81 (2021)annual/2022- 117-262 (2022)congress/117- 117-262 (2022, end of 117th Congress)annual/2024- 118-158 (2024)congress/118- 118-158 (2024, end of 118th Congress)annual/2025- 119-73 (2025, current latest)
Limitations
Coverage starts in 2013 as OLRC XML release points are only available from the 113th Congress onward. The repository contains codified law only (the consolidated United States Code), not individual bills or public laws in directive format. Not all titles are positive law - some titles are "evidence of law" rather than the legal text itself. Appendix titles (5A, 11a, 18a, 28a, 50A) are not yet included, and 6 sections across titles 5, 10, 25, 28, 38, and 40 have duplicate section numbers in the source XML.
📖 Read the full source: HN LLM Tools
👀 See Also

AnyClaw: Ubuntu 24.04 with Android hardware access and AI agent for terminal development
AnyClaw provides a full Ubuntu 24.04 environment running in proot on Android with direct access to Android hardware APIs from the terminal, including camera, GPS, battery, and sensors via bash commands and Java execution. It includes an AI coding agent that can orchestrate these tools and a web UI accessible from any browser on the same network.

bad-ass-mcp: Free, Open Source MCP for Native Desktop GUI Control via Accessibility API
bad-ass-mcp is an open source MCP server that lets Claude and other AI agents control macOS, Windows, and Linux desktops using the native accessibility layer — no screenshots, no look-move-look loops. Free alternative to Computer Use, Operator, or UiPath.

Open-source CLI uses Claude Haiku to automate Xero expense auditing
A developer has released an open-source Python CLI tool that uses Claude Haiku 4.5 to automate Xero expense auditing. The tool follows a 'deterministic code first, then AI to fill in the gaps' approach, keeping costs to a few cents per audit run.

AutoBe: How Weak Local LLMs Fixed an AI Backend Generator's Architecture
AutoBe is an open-source AI agent that generates complete backend apps using TypeScript, NestJS, and Prisma. The team discovered their initial 100% compilation success produced unmaintainable code, then rebuilt with modular generation—crashing success to 40%—and used weak local LLMs like qwen3-30b-a3b-thinking to debug schema ambiguities.