Documentation

Introduction

Rezos is an AI operations platform that wraps any LLM with a cryptographic verification layer. Every file operation an agent claims to perform is validated against the actual filesystem before being accepted.

What is Rezos?

Modern AI coding assistants are powerful but operate on trust. When a model says “I've updated your file,” you typically have no way to verify that claim short of manually checking the diff. This creates invisible risk — phantom edits that look complete in the conversation but never actually happened, or worse, edits that happened but not as described.

Rezos eliminates this by wrapping every LLM call with Strict Write Discipline (SWD) — a protocol that takes SHA-256 snapshots before and after every file operation, compares the claimed changes against actual disk state, and only seals a tamper-evident receipt when reality matches the claim.

If a mismatch is detected, Rezos gives the model up to 2 correction turns to reconcile the difference before yielding control back to the human.

Key concepts

Strict Write Discipline (SWD)
The core verification protocol. Pre/post SHA-256 snapshots verify every model file claim. Mismatch triggers correction turns.
SWD Receipt
A tamper-evident log sealed after a verified session. Contains touched files, hashes, provider identity, cost, and verification result.
Drift
A discrepancy between what an agent claimed to write and what actually landed on disk. Rezos surfaces drift in real time.
Skill Packs
SKILL.md files (project-local or user-global) that encode rules for the AI. Like .cursorrules but verified and composable.
Memory
A self-healing SQLite FTS5 index rebuilt from MEMORY.md as authority. Persists context across sessions without hallucination risk.

Installation

Terminal
# Install globally via npm
$ npm install -g rezos-router
# Set your API key
$ export ANTHROPIC_API_KEY=“sk-ant-...”
# Initialize a project
$ rezos init
# Start a verified session
$ rezos chat

Next steps