Introduction

Tenbri is an open-source, tamper-evident record of what AI agents decided: who acted, under which rule and policy version, from which inputs, and who approved. It is written in Rust, stores plain files on a local disk or an object store, and verifies offline against a trust root the verifier controls.

It works on its own, with any agent framework, from Rust, Python, Node, the browser (WASM), the tenbri CLI or the tenbrid REST and gRPC API.

The problem

Companies now give agents real authority: refunds, payments, account changes, production deploys, security responses. Afterwards they cannot prove what an agent did, what it saw, which rule allowed it, or who signed off.

A customer disputes a refusal. Support can show the reply, but not which policy version, contract clause and permission produced it.
An incident happens. Engineers rebuild the story from logs in five services, each editable by the people being investigated.
A policy changes. Nobody can list the decisions made under the old wording.
An auditor asks for evidence. Screenshots and CSV exports prove nothing about integrity.
A prompt injection makes an agent act. There is no record of what the agent read.

Non-goals

Multi-writer consensus, a blockchain or token, a general-purpose database, storing raw telemetry or full prompts by default, judging whether a decision was correct, telemetry of any kind, and a hosted service.