Skip to content

WarrantOS

Every factual claim in an AI-assisted document ships with a source, a [CITE NEEDED] tag, or a logged BLOCK. No warrant, no ship.

claude-provenance on GitHub and as the legacy Claude Code plugin; warrantos on PyPI and the CLI.

Install

pipx install warrantos      # isolated CLI install
uvx warrantos demo          # zero-install trial run
pip install warrantos       # plain pip works too

New here? Start with the Quickstart for a five-minute tour, or Overview for what ships in this repository.

Three ways in

Lint your docs for AI slop. Scan for scaffold residue and unsourced claims before a draft goes anywhere. Rules live in one place so a slop finding and a check violation are always explained the same way. See the Specification.

Gate your agent. A stdio MCP server and a Claude Code Stop hook check what the model wrote before the turn ends: stdlib only, no network, never breaks the session. See Plugin and the MCP config reference.

Build an audit trail for governance. Seal a checked run into a portable .warrant bundle that a third party verifies offline, with a zero-backend browser verifier for readers who have only the file. See Verification and the Compliance mapping to ISO/IEC 42001 and the NIST AI RMF.

What this does not do

The claim detector is a heuristic and will produce false positives and false negatives. Offline verification checks token overlap, not meaning. It does not replace human review and does not claim to. Read the full list in Limitations.

Where to go next

  • Status: the per-layer build-state dashboard
  • Distribution: install surfaces and packaging
  • Reference section in the sidebar: deeper design notes, the roadmap, and the multi-agent review record

Source, issues, and releases: github.com/jvega017/warrantos