Proof Engine

Proof Engine

[](https://proofengine.info/static/install-claude-desktop.html)

[](https://opensource.org/licenses/MIT) [](https://agentskills.io) [](https://docs.anthropic.com/en/docs/agents-and-tools/claude-code/plugins) [](https://cursor.com/docs/plugins)

An AI agent skill that verifies claims through code and live sources — not by asking the LLM to check itself. Every fact is either computed by Python code anyone can re-run or backed by a specific source, URL, and exact quote. The LLM never asserts a fact on its own authority.

Uses the open [Agent Skills](https://agentskills.io) standard. Works with Claude Desktop, Claude Cowork, Claude Code, Codex CLI, Cursor, Windsurf, Manus, ChatGPT, and any other compatible tool.

Why This Exists

LLMs hallucinate facts — and they hallucinate the checks on those facts. Ask an LLM to verify its own claim and it runs the same error-prone process again. The check is circular.

Proof Engine breaks the circle by routing every claim through a gate the LLM can't fake:

• Computations are Python — Python doesn't hallucinate. If the LLM sets up the wrong calculation, the result is wrong in a visible, re-runnable way.
• Citations are fetched and matched — the script hits the URL and searches for the quoted text on the live page. A fabricated citation fails the match; a partial match downgrades the verdict so the gap is visible.
• 9 hardening rules close the remaining escape routes — don't hand-type val